Electric Vehicle Charging IoT

Electric Vehicle Charging IoT — Topics for IoT Students

The Role of Generative Artificial Intelligence in

Internet of Electric Vehicles Hanwen Zhang , Student Member, IEEE, Dusit Niyato , Fellow, IEEE, Wei Zhang , Member, IEEE, Changyuan Zhao , Hongyang Du , Abbas Jamalipour , Fellow, IEEE, Sumei Sun , Fellow, IEEE, Yiyang Pei , Senior Member, IEEE

Abstract—With the advancements of generative artificial I. I NTRODUCTION

arXiv:2409.15750v3 [cs.LG] 14 Nov 2024

intelligence (GenAI) models, their capabilities are expanding significantly beyond content generation and the models are increasingly being used across diverse applications. Particularly, GenAI shows great potential in addressing challenges in the E LECTRIC mobility is the future. This is evidenced by the visions and policies of different nations and companies for achieving sustainable mobility.

Electric vehicles (EVs) are electric vehicle (EV) ecosystem ranging from charging manage- gaining popularity rapidly. In the US, the sales of EV reached ment to cyber-attack prevention. In this paper, we specifically 1.2 million just in one quarter in 2023, with nearly 8% market consider Internet of electric vehicles (IoEV) and we categorize share [1].

The numbers are even bigger in China. In 2023, EV’s GenAI for IoEV into four different layers namely, EV’s battery layer, individual EV layer, smart grid layer, and security layer. market share was 34% [2]. The trend is not much different in We introduce various GenAI techniques used in each layer of other nations, e.g., the sales of traditional cars with internal IoEV applications.

Subsequently, public datasets available for combustion engines (ICEs) will be phased out by 2030 and training the GenAI models are summarized. Finally, we provide 100% cars will be clean energy based after 2040 in Singapore. recommendations for future directions. This survey not only Altogether, the global market of EVs reached 392.4 billion USD categorizes the applications of GenAI in IoEV across different in 2023 with a predicted compound annual growth rate of 13.9% layers but also serves as a valuable resource for researchers from 2024 to 2032 [3].

With more and more EVs on the road, and practitioners by highlighting the design and implementation they naturally form a IoEV [4]. It shares a similar spectrum of challenges within each layer. Furthermore, it provides a roadmap technologies to Internet of things (IoT) and offers new features for future research directions, enabling the development of more robust and efficient IoEV systems through the integration of such as connectivity, grid services, predictive maintenance, and advanced GenAI techniques. traffic management.

IoEV could be regarded as a specialized subset of IoT where EVs, their charging infrastructure, and Index Terms—Generative artificial intelligence, Internet of smart grid interact seamlessly [5]. This creates a network of electric vehicles, scheduling, forecasting, scenarios generation interconnected devices/systems, leading to various applications, e.g., EV routing problem in IoEV [6], smart EV charging station scheduling [7], EV’s battery life prediction [8], and Manuscript received January 1, 2000; revised January 1, 2000.

This blockchain-based bidirectional energy trading between EVs research/project is supported by the National Research Foundation, Sin- gapore and Infocomm Media Development Authority under its Future and charging stations [9]. Communications Research & Development Programme (Grant FCP-SIT-TG- Same as most new technologies, EV and IoEV have their 2022-007), A*STAR under its MTC Programmatic (Award M23L9b0052), problems and challenges.

Making electricity the main source MTC Individual Research Grants (IRG) (Award M23M6c0113), the Ministry of power brings constraints related to electricity at the same of Education, Singapore, under the Academic Research Tier 1 Grant (Grant ID: GMS 693), and SIT’s Ignition Grant (STEM) (Grant ID: IG (S) 2/2023 time. The electricity is stored in batteries, which have charging – 792). (Corresponding author: Wei Zhang) speed and capacity limits, can degrade over time, and may Hanwen Zhang is with both the College of Computing and Data catch fire occasionally.

Those constraints affect EV operations Science, Nanyang Technological University, Singapore 639798, and the in various aspects, e.g., charging scheduling and battery health Information and Communications Technology Cluster, Singapore Institute monitoring. Furthermore, the impact goes beyond individual of Technology, Singapore 138683 (e-mail: hanwen001@e.ntu.edu.sg and hanwen.zhang@singaporetech.edu.sg) EV for IoEV with increased coordination and intelligence Wei Zhang, Sumei Sun, and Yiyang Pei are with the Informa- demand.

One impact is on the power grid, where a significant tion and Communications Technology Cluster, Singapore Institute amount of new electricity load from EVs shall not stress the of Technology, Singapore 138683 (e-mail: {wei.zhang, sumei.sun, grid much and hurt the grid’s stability. This requires grid-level yiyang.pei}@singaporetech.edu.sg). intelligence such as supply-demand forecasting and matching Dusit Niyato and Changyuan Zhao are with the College of Computing and Data Science, Nanyang Technological University, Singapore 639798 with the support of IoEV scheduling and vehicle-to-grid (V2G) (e-mail: dniyato@ntu.edu.sg and zhao0441@e.ntu.edu.sg). services.

Hongyang Du is with the Department of Electrical and Electronic The problems and challenges have drawn attention from Engineering, University of Hong Kong, Pok Fu Lam, Hong Kong (e-mail: the research communities and industries. Existing efforts can duhy@eee.hku.hk). Abbas Jamalipour is with the School of Electrical and Com- be clustered into application level and technology level.

The puter Engineering, the University of Sydney, Australia (e-mail: ab- counterparts of EV and IoEV are traditional cars with ICEs. bas.jamalipour@sydney.edu.au). IoEV is a subset of Internet of vehicle (IoV), with high

relevance. Research studies for these applications have the TABLE I potential of being transferred to IoEV, but this comes with L IST OF ABBREVIATIONS . certain constraints and potentially substantial costs for some tasks due to the new challenges of IoEV. For example, charging Abbreviation Description AE Autoencoder typically takes hours and cannot be modeled as an instant event, AM DRL with general attention model and electricity prices may change hourly or more frequently.

As ANN Artificial neural network such, the EV charging management system shall schedule and AutoML Automated machine learning predict the EV loads to minimize cost and avoid overloading BMS Battery management system the power grid, and charging station installation and operation CAN Controller area network shall be optimized to promote charging service availability. CNN Convolutional neural network Technology-wise, there are solutions specialized to IoEV, DBN Deep belief network DDIM Denoising diffusion implicit model with or without using machine learning (ML).

Non-ML DDPM Denoising diffusion probabilistic model solutions often use model-based statistical methods for various DNN Deep neural network IoEV related tasks, e.g., user behavior understanding [10] and DoS Denial of service charging scheduling [11]. A common problem of these methods DRL Deep reinforcement learning is that the real system dynamics are modeled in a highly ESS Energy storage system summarized way, e.g., mean and variance, and formulated with EV Electric vehicle EVRP Electric vehicle routing problem simplification to make the optimization tractable.

As a result, FDIAs False data injection attacks they are often impractical and fail to offer sufficient accuracy FGSM Fast gradient sign method and effectiveness in modeling and charging scheduling. ML- FNN Feed-forward neural network based solutions are becoming the trend. ML models learn from GAE Graph autoencoder data and continuously improve with more data available.

For GenAI Generative artificial intelligence example, the long short-term memory (LSTM) is often used in GAN Generative adversarial network GDM Generative diffusion model the prediction of EV load [12] and voltage changes of battery GMM Gaussian mixture model [13] because of its superior capability of handling long-term GPR Gaussian process regression dependencies in sequential data, yet the prediction accuracy EMS Energy management system decreases when uncertainty in the real systems increases.

IoEV Internet of electric vehicles In this paper, we propose to use GenAI to advance IoEV IoT Internet of things technologies. GenAI techniques applied in IoEV mirror the IoV Internet of vehicle KNN k-nearest neighbor same advancements in IoT where a large amount of data LLMs Large language models from various sources are analyzed and utilized to enhance LSTM Long short-term memory efficiency, safety, and user experience.

The integration of GenAI MADRL Multi-agent deep reinforcement learning within IoEV not only enhances its specific applications but also MAE Mean absolute error contributes to the overarching goals of IoT by enabling smarter, MARL Multi-agent reinforcement learning more autonomous, and connected systems. Moreover, GenAI ML Machine learning NLP Natural language processing has the potential to address the challenges mentioned above and PPO Proximal policy optimization case studies for certain IoEV tasks are available, e.g., charging PV Photovoltaic demand forecasting [14], [15] and data augmentation [16], [17].

RL Reinforcement learning We aim to go beyond case studies and provide a comprehensive RNN Recurrent neural network discussion of GenAI’s roles in the IoEV ecosystem. Specifically, SAC Soft actor-critic we structure the system into four layers as shown in Fig. 1. The SoC State of charge SoH State of health bottom layer is the battery, which is a critical component of SVM Support vector machine an EV and brings many new constraints such as long charging VAE Variational autoencoder time and battery degradation to IoEV compared to traditional VRP Vehicle routing problem IoV.

Next to the battery layer is the EV layer, where EVs are considered individually for various aspects such as EV charging behaviors and load, as well as an EV routing problem. Then, we consider the existence of many EVs that collectively share a few of various GenAI models. We also highlight their roles charging stations connected to the smart grid, which becomes in solving various IoEV problems such as data scarcity the third layer.

This layer features the aggregated demand and charging load prediction. aims to achieve the optimal charging scheduling of many EVs. • We systematically categorize GenAI-enabled IoEV ap- Finally, we provide an investigation of the security layer that plications into four distinct layers for battery, individual is located vertically across the above three layers. For reader’s EV, the grid, and security. We describe each layer with convenience, we also present a list of common abbreviations the specific GenAI techniques for their respective IoEV for reference in Table I.

In summary, we make the following applications, and provide a holistic view of how GenAI main contributions in this paper. can be integrated within the IoEV ecosystem. • We present a detailed survey of the latest GenAI techniques • We provide a summary of publicly available datasets across all layers of IoEV, including an in-depth exploration for training GenAI models within the IoEV context. It

Layer 4: Security Layer 3: Smart Grid with EV

Problems to be addressed: Data quantity and quality issues, e.g., EV charging load profiles Predicting price and load for Synthetic Price EV Load EV Charging optimal EV charging schedules Data Forecast Forecast Schedules Optimizing EV charging schedules

Layer 2: Individual EV

Problems to be addressed: Investigation of Problems to be addressed: various attacks Data scarcity, e.g., EV charging behaviors Detection of both (arriaval and departure time, and SoC cyber and physical level) Synthetic EV Charging EV Routing attacks Predicting EV charging behaviors and load Data Behaviors Problem Solving for optimal EV routing problem

Layer 1: EV's Battery

Cyber Attacks Problems to be addressed: Data scarcity, e.g., battery's voltage, current, SoC, and degradation status Detecting anomaly earlier before failure Synthetic Anomaly SoC & SoH Physical Attacks Monitoring battery status via SoC and SoH Data Detection Estimations estimations

Fig. 1. GenAI for IoEV applications can be categorized into four layers: Layer 1’s problems are anomaly detection, SoC estimation, and SoH evaluation. Layer 2 primarily focuses on data augmentation in EV charging behaviors, prediction of an EV load at home, and solving the optimal EV routing problem.

Layer 3 concentrates on: Forecasting and augmenting the EV charging load profiles; Optimizing EV charging schedules based on the constraints from either EV charging stations or the smart grid; Predicting the electricity price is necessary for EV charging station operators. Layer 4 studies various attacks which may be harmful to the EV and the charging system. Both cyber and physical attacks need to be studied and detected such as adversarial attacks, false data injection attacks, denial of service attacks, fuzzy attacks, and impersonation attacks.

facilitates future research on GenAI for IoEV applications. station operators (CSOs) and smart home owners to optimally • We identify and discuss the critical challenges and gaps schedule the charging/discharging of connected EVs. In this in the existing GenAI applications in IoEV. After that, case, the DSO sends requests to the contracted CSOs and smart we suggestion future research directions aiming at solving home owners to increase/decrease the energy consumption in a existing challenges and providing new opportunities. certain period of time.

Then, the CSOs and smart home owners • We bridge the gap between multiple disciplines including: respond to the DSO’s request and optimally schedule the EV computer science, electrical engineering, and transporta- charging/discharging, considering photovoltaic (PV) generation tion. This comprehensive survey could serve as a valuable and an energy storage system (ESS) for achieving different resource for a wide audience. objectives such as charging cost minimization.

To participate The subsequent sections of this paper are organized as in the day-ahead energy market, CSOs often need to forecast follows. In Section II, the basic concepts of GenAI and the electricity price and EV load for optimal EV scheduling in EV charging system are initially conducted to give a brief advance. Moreover, EV charging behaviors such as arrival time, background.

Following this, a comprehensive review of GenAI departure time, charging duration, and unplugging time can techniques as they pertain to IoEV is undertaken in Section III. influence EV load forecasting. From the EV users’ perspective, Subsequently, the available public datasets are summarised in they may not only be interested in saving charging costs through Section IV. Next, the future directions are recommended in smart home systems but also need to understand EV’s battery Section V.

Lastly, a conclusion is presented in Section VI. status. Two important aspects are state of charge (SoC) and state of health (SoH). The former refers to the remaining quantity II.

BACKGROUND of electricity available in the EV battery, which implies the remaining range. The latter indicates the aging status of the In this section, we introduce the concept of EV and IoEV, battery for making decisions about battery maintenance and the basics of GenAI models, as well as a brief introduction of retirement [18]. Furthermore, an IoEV ecosystem, is formed GenAI’s industry adoption. by connecting DSO, CSOs, and EV users, and the optimal EV scheduling can be achieved through coordinated efforts among A.

Electric Vehicle (EV) and Internet of EV (IoEV) the IoEV components. The approach ensures that the concerns of a DSO, CSOs, and EV users are well considered and Fig. 2 shows the concept of an EV charging system in addressed based on their respective interests and constraints. an electrical distribution network. The network consists of charging stations, residential buildings (e.g., smart homes), load, and distributed resources (e.g., renewable energies, and grid- B.

Basics of Generative Artificial Intelligence (GenAI) scale energy storage systems). A distribution system operator GenAI is proposed based on traditional ML, which are (DSO) being one of the grid operators, manages the electrical discriminative models that learn the probability distribution distribution network. The DSO can coordinate with charging p(y|x) in Bayes’ theorem for input x and output y.

The

Charging Station 2 Electrical EV routing [26], and EV charging load forecasting [27]. The

Transformer Distributed Electric unique strength of capturing long-range dependencies makes Resources Vehicle (EV) Transformer suitable for parallel computing, scalable according EV Charger to the different tasks, and transferable with a pretrained Smart Photovoltaic (PV) model. However, the Transformer is not perfect. For example, Grid Energy Storage the Transformer used in LLMs often requires significant System (ESS) computational and memory resources during the training.

The Distribution DSO System performance of the Transformer may be compromised when Operator (DSO) there is only a small amount of dataset available for training. Charging Station 2) Generative Adversarial Network (GAN): The GAN Smart Operator (CSO) model [28] consists of two deep neural network (DNN): a Home Load EV User generator and a discriminator. Both networks engage in an EV's Battery Charging Station 1 adversarial training process; one network generates new data while the other assesses whether the data is real or fake.

Fig. 3 Fig. 2. Concept of EV charging system in an electrical distribution (b) shows the principle of GAN. Assume that the data xi is network: the network consists of charging stations, smart home, distributed extracted from the data distribution pdata (x), with the goal resources, and load.

Distribution system operator (DSO) and charging of sampling it according to pdata . The sample z, the latent station operators (CSOs) manage the distribution network and the charging stations respectively. Smart homes and charging stations with PV and ESS variable, drawn from the simple prior p(z) is fed into the can coordinate with DSO to ensure a stable and robust smart grid operation. generator network.

Then, the generated sample is drawn from the data distribution of the generator pg . Subsequently, the joint training of the generator and discriminator is carried out until discriminative models categorize the data space into different pg converges to pdata , i.e., pg ≈ pdata . In the training process, classes by learning the decision boundaries.

They often focus the generator network is trained to “deceive” the discriminator on distinguishing between different classes or outcomes. In that concurrently learns to classify whether the generated data the context of computer vision applications, the discriminative is real or fake with generator and discriminator loss functions. models are incapable of processing unknown inputs and it is From the mathematical point of view, it is similar to a two- required to provide label distributions for every image.

As player minimax game with an objective function. such, traditional ML models are mainly used for classification, GAN was first proposed in 2014 [28]. Some evolved regression, clustering, etc. versions are deep convolutional GAN (DCGAN) [29] in 2016, GenAIs are different from traditional ML and they learn the Progressive GAN [30] and Wasserstein GAN [31] in 2017, probability distribution of data p(x) for unconditional genera- StyleGAN [32] in 2019 as well as its adoption in semantics tive models or p(x|y) for conditional generative models.

This communication [33], [34], and MaskGAN [35] in 2020. Besides allows GenAIs to understand the underlying data distribution its applications in computer vision, GAN is also extended and generate new samples from the distribution. The generated to other domains, e.g., data augmentation for battery’s SoC data/contents could be statistically similar to the input data estimation [36], data augmentation for EV charging behaviors and the similarity is useful for data augmentation, simulation, [17], data augmentation for EV charging load profile [37], and creative tasks.

Overall, the generative nature of GenAI is EV load forecasting or changing scenarios generation [38], useful for developing more dynamic and innovative solutions as well as generation and detection of adversarial attacks across various domains. [39]. Moreover, the idea of GAN combined with imitation 1) Transformer: The Transformer architecture was first learning formed a new term called generative adversarial introduced by the influential article “Attention is all you need” imitation learning (GAIL) [40] which is particularly use- [19] which utilizes a self-attention mechanism to capture long- ful in reinforcement learning (RL) when defining a reward range dependencies without relying on sequential processing. function is explicitly difficult.

Besides, a variant of GAN Fig. 3 (a) shows a single-layer Transformer that includes is generative adversarial imputation network (GAIN) [41], a typical self-attention module and feedforward layers with which is particularly useful for missing data imputation and residual connections where each layer begins with the applica- accordingly improves performance. This is important for IoEV tion of self-attention. The resulting output from the attention applications that involve incomplete datasets, e.g., with missing mechanism is then processed by feedforward layers, where the entries in charging behavior logs or irregularities in electricity same feedforward weight matrices are used independently for consumption records.

While GAN is capable of learning each position. After processing through the first feedforward complex and high-dimensional data distribution as well as layer, a nonlinear activation function, e.g., ReLU, is applied. generating high-quality data, there are challenges, such as The Transformer structure is widely used in large language balancing generator and discriminator, avoiding mode collapse, models (LLMs) [20], image processing [21], automatic speech and accelerating convergence. recognition [22], visual question answering [23], sentiment 3) Autoencoder (AE) and Variational Autoencoder (VAE): analysis [24], etc.

Besides the traditional ML applications such An AE is an unsupervised approach that extracts feature vectors as audio, computer vision, and natural language processing from raw data x without labeled examples. It consists of the (NLP) applications, Transformer is also extended to other encoder and decoder during the training as shown in the left domains such as the anomaly detection of the EV battery [25], side of Fig. 3 (c).

The encoder learns the useful information,

(a) Transformer (b) Generative Adversarial Network (GAN)

Residual Connection Residual Connection Generator

Data Latent Space Real Generated Generator Loss

+

+ Sample Sample

Input Self-Attention Feedforward Output Discriminator

Vectors Vectors Discriminator Real/Fake? Transformer Layer Loss

(c) Autoencoder (AE) and Variational Autoencoder (VAE) (d) Generative Diffusion Model (GDM) Encoder (Forward Diffusion) Loss Function Loss Function

Label ... ...

Input

Encoder Decoder Encoder Classifier ... ... Input Reconstructed Input Predicted Data Features Input Data Data Features Label Output

Training After Training Decoder (Reverse Diffusion)

Fig. 3. Concept of basic GenAI models — Transformer, GAN, AE, VAE, and GDM: (a) shows a single-layer Transformer process where the output vectors are achieved by passing the input vectors through self-attention and feedforward layers with residual connections; (b) illustrates principles of GAN where the generator competes with the discriminator by producing increasingly realistic samples to “fool” the discriminator, while the discriminator attempts to differentiate between real and fake data; (c) depicts principles of AE: the left side shows the training process of AE, while the right side depicts its usage once the model is completely trained; VAE is a subcategory of AE, but it differs slightly in that VAE uses the mean and diagonal covariance to generate samples in both the encoder and decoder; (d) displays the processes of GDM consisting of forward diffusion and reverse diffusion.

i.e., features Z, from the input raw data x. The encoder could optimized solely through gradient-based methods. The concept be Sigmoid, fully connected, or ReLU convolutional neural of VAE is similar to AE, as it represents a specific subset of network (CNN).

After that, the decoder utilizes the learned AE. The encoder and decoder of the original AE are modified features to reconstruct the input data x̂. The decoder could be in the VAE where the sampling processes are achieved using Sigmoid, fully connected, or ReLU CNN (up-convolution or means and diagonal covariances.

The VAE has been used for transposed convolution). The loss function of the training could common ML applications such as facial expression editing be the L2 distance between input and reconstructed data. After [48], future forecasting from static images [49], and point the training, the decoder is removed and the trained encoder cloud completion [50].

Its usage has also been extended to is useful for the downstream tasks as shown on the right side IoEV applications, e.g., anomaly detection for EV’s battery of Fig. 3 (c). For example, a supervised classification model [51] and data augmentation for EV load profiles [14]. Generally, can be initialized using the encoder which is often fine-tuned VAE has better generative capabilities compared to vanilla AE, jointly with the classifier and a task-specific loss function. and can generate new samples similar to the training dataset.

The capabilities make VAE suitable for data augmentation and AE can be used as the context encoder in semantic image synthesis, and useful for the clustering and interpolation inpainting tasks [42], the temporal context encoder for video tasks. However, using a Gaussian prior and a reconstruction applications [43], and representation learning [44]. Besides, loss function, VAE may not be able to capture fine details well, AE is useful in EV related applications, e.g., detection of false e.g., in images with blurry outputs.

Moreover, VAE may suffer data injection attacks (FDIAs) that may pose a threat to the EV from mode collapse where the model generates a few types of charging process [45]; cyber and physical anomaly detection for outputs despite having diverse data. Its generative capabilities abnormal behaviors within EV charging stations [46]; detection can be also limited by the latent space Gaussian assumption. of denial of service (DoS), fuzzy, and impersonation attacks to the controller area network (CAN) protocol communications of 4) Generative Diffusion Model (GDM): Fig. 3 (d) shows EVs connected to EV charging system [47].

AE offers several the principle of the GDM. In its forward diffusion process, benefits. It can effectively reduce the dimensionality of the data i.e., encoder, the model transforms x through a sequence and learn the compact representations, which enables it to be of latent variables z1 , . . . , zT .

The procedure is predefined useful for data preprocessing and feature extraction. It allows and progressively blends the data with noise until only noise unsupervised learning since no labeled data is required for persists at zT . Given a sufficient number of steps, both the training, and is suitable for identifying anomalies by learning conditional distribution q(zT |x) and the marginal distribution to reconstruct normal data well.

However, vanilla AE lacks q(zT ) of the final latent variable converge to the standard generative capabilities compared to VAE and GAN, and cannot normal distribution. All the learned parameters are included in capture complex data distributions effectively. the decoder as predefined. In the reverse diffusion process, the The VAE is a type of directed model that relies on data is processed through the latent variables by the decoder approximate inference learned during training and can be which is trained to eliminate noise progressively at each stage.

The backward mapping between each pair of adjacent latent invented for traditional ML tasks such as computer vision and variables zt and zt−1 is achieved through the training of a NLP and gradually adapted for electric mobility applications. sequence of networks. Each network is guided by the loss For EV batteries, GenAI helps estimate battery capacity and function to perform an inversion of its associated encoder step. health and detect potential anomalies.

For EVs, GenAI can be Following the training process, the new examples are created used to understand charging behavior, energy management, and by sampling vectors of noise zT and then processing these routing. With many EVs forming an IoEV, data augmentation through the decoder. and large-scale charging scheduling can be supported by GenAI. The diffusion model has been widely utilized in image Furthermore, GenAI’s role in detecting cyber and physical applications, e.g., the denoising diffusion probabilistic model attacks on IoEV components can be explored.

Overall, adapting (DDPM) [52] for generating high-quality image samples foundational GenAI models for various IoEV aspects from without adversarial training, denoising diffusion implicit model batteries to security requires careful consideration of GenAI’s (DDIM) [53] for improving the sampling speed of DDPM, respective strengths and limitations. stable diffusion [54] for generating images from text, and ControlNet [55] allowing model being trained with a small dataset of image pairs.

Furthermore, the diffusion model has C. GenAI for Industry been extended for other applications, e.g., network optimization With the basics of GenAI, we present adoption of AI and [56], generating optimal pricing strategies [57], repairing GenAI in EV industry as well as other domains. and enhancing extracted signal features in wireless sensing 1) EV Industry: GenAI has not yet been specifically [58], estimating the signal direction of arrival in near-field presented in the EV industry but AI has been among the scenarios [59], estimating battery’s SoH [60], and generating strategic focuses of the big EV players and we introduce EV charging scenarios [61].

GDM can produce high-resolution some latest progress from two major players. AI has been and realistic samples, achieving similar performance of GAN adopted in various aspects of Tesla’s business, e.g., EV generated data or even better, e.g., outperforming the traditional manufacturing and autonomous driving. The large-scale AI Gaussian mixture model (GMM) model by 91% for charging adoption is enabled by Tesla’s Dojo supercomputer which offers scenarios generation.

Compared to GAN, GDM shows better abundant computing resources and realizes computational- training stability in versatile applications, e.g., optimization intensive tasks such as high-throughput EV video processing. [56], [57]. Moreover, the clear and iterative process of refining The company also owns an overarching aspiration to develop the generated data from noise to coherent output makes the artificial general intelligence (AGI).

BYD has introduced its generation process of GDM more interpretable. Nevertheless, XUANJI Architecture, an intelligent vehicular framework, the iterative nature of the process can be the drawback of GDM. integrating electrification with intelligent functionalities and It requires a significant amount of computational resources for functioning as the EV’s cognitive core. EV’s internal and the training and inference of GDM.

This nature also slows external environments are monitored in real-time and the down the sampling process of GDM compared to the VAE and collected information is used to make decisions about the EV’s GAN models that can generate samples in a single pass. When operation to improve safety and comfort. The industry favors designing and tuning the GDM, the noise schedules and model the integration of AI and physical systems and we foresee an architecture have to be carefully considered.

This increases increasing adoption of GenAI in the EV industry. the complexity of training well-performed models. Moreover, 2) Other Industry Sectors: Besides EV and the transporta- GDM lacks controllability in specific attributes or features. tion sector, GenAI has found widespread adoption in other 5) GenAI Development and Deployment: Three important industry sectors and we introduce a few sectors as follows. stages are training, fine-tuning, and deploying the above- For business and finance, a survey [62] is available with mentioned GenAI models in practice.

First, GenAI models a description of GenAI’s practical applications and cutting- are trained on large and diverse datasets to learn broad edge tools in the sector. The survey [63] is about GenAI’s patterns and general features. Foundation GenAI models are applications, advantages, and obstacles for the healthcare sector. produced in this stage and serve as the backbone of various The education sector is witnessing GenAI’s significant impact applications.

However, such models are general-purposed, so a and the authors in [64] discuss GenAI’s ability to boost learner fine-tuning stage is needed to customize the models for specific engagement and motivation, emphasizing the need for ethical applications, such as EV charging, load forecasting, and route guidelines and human oversight as well as GenAI’s impact on optimization. With the support of the backbone and given critical thinking.

Compared to different industry sectors, the the realistic constraints, the stage involves small-scale datasets IoEV industry has various unique features, e.g., battery and only which, however, shall be domain-specific. Finally, the fine- grid integration. The features require GenAI algorithms to be tuned models are deployed in practice for real-time inference customized and specialized for performance maximization. which is significantly less computing intensive compared to training and fine-tuning models.

Overall, the three stages, with III. T ECHNICAL R EVIEWS : G ENERATIVE A RTIFICIAL different resource demands and objectives, orchestrate GenAI I NTELLIGENCE (G ENAI) FOR I NTERNET OF E LECTRIC development and deployment. V EHICLES (I O EV) 6) Summary: Transformer, GAN, AE, VAE, and GDM are foundational models of GenAI which have demonstrated their In this section, we provide an overview and discussions of versatility across a wide range of applications.

They were first GenAI’s application in different layers of IoEV.

A. Layer 1: Electric Vehicle (EV)’s Battery recurrent unit (GRU) and a VAE, where the former captures The battery is the core component of an EV. It powers MVTS and the latter reconstructs the input samples.

Worth an electric motor and has a direct impact on EV’s range, mentioning that the paper is based on an EV operation dataset performance, and efficiency. It is also the most expensive part from the National Service and Management Center for Electric of most EVs, due to the fact of which, battery’s operating Vehicles (NSMC-EV) in Beijing [74]. The dataset includes 13- condition and longevity play a crucial role in the overall user dimensional time series signals such as data acquisition time, experience and sustainability of the EVs.

We specifically would vehicle speed and state, charging state, voltage, current, mileage like to survey three important aspects of batteries, including accumulated, SoC, temperature, insulation resistance, and anomaly detection [25], [51], SoC [36], [65], [66], and SoH DC–DC state. GRU-VAE shows an improvement in anomaly [60]. For anomaly detection, various detection algorithms can detection compared to the AE-based models, achieving a 25% be integrated with the battery management system (BMS) for increase in F1-score [51].

Nevertheless, the model may need proactive battery maintenance before any potential failures to be updated, given the changed data patterns over time. cause hazardous battery damage. The SoC and SoH indicate c) Transformer-based Approach: Anomaly detection the battery’s short-term energy capacity and long-term health can also be transformer-based [25]. The authors in [25] devel- condition, respectively.

Specifically, SoC measures the stored oped BERTtery, a transformer-based model for battery fault energy relative to the maximum capacity and SoH reflects the diagnosis and failure prognosis. BERTtery is able to capture battery’s maximum capacity which degrades over time. Both early-warning signals across multiple spatial–temporal scales aspects are influenced by different factors such as temperature, in various operational conditions, predict battery system fault charging voltage, and cycling history [18]. evolution using onboard sensor data, and avoid faults leading 1) Anomaly Detection: Despite continuous technological to thermal runaways [75].

The model is based on a dataset with progress in the past years, battery safety remains a big concern various battery faults and failures, e.g., internal short circuits, for EV owners and customers. One of the most hazardous issues lithium plating faults, overcharging/overdischarging, abnormal is battery fire, the occurrence of which has raised debates and self-discharge, abnormal capacity degradation, abnormal volt- doubts about battery safety and highlights the necessity of age fluctuations, abnormal temperature behaviors, electrolyte early anomaly detection to prevent potential safety breaches leakages, cell balancing issues, and thermal runaways [76]. and irreversible damage [67].

Specifically, the model’s input includes the time series of a) Traditional Machine Learning (ML) Approaches: voltage, current, and temperature, sampled every 10 seconds The early effort of battery anomaly detection and diagnosis is from real-world EV operations. The output of the model is based on traditional ML algorithms, e.g., the random forest [68], the predicted safety labels. For future work, the generalization multiclass relevance vector machine [69], and finite-element- ability of the proposed model could be improved by considering based models [70].

These ML algorithms are generally simple different battery types and operational conditions, to enhance to implement but the performance suffers when the raw data is the early warning capabilities with reduced false alarms. used directly. Domain knowledge complements the capability 2) State of Charge (SoC) Estimation: SoC indicates how of the algorithms by guiding the extraction of domain-specific much battery energy remains and directly affects the EV’s and useful features as the ML input, with improved correlation range [77].

Thus EV owners monitor SoC for trip planning, with battery anomalies. The advancements of deep learning charging scheduling, and battery usage optimization. offer new methods for battery anomaly detection. Among the a) Traditional ML-based Approaches: Many ML algo- methods, LSTM should be the most widely used architecture rithms have been adopted for SoC modeling, e.g., the random [71], [72], which is capable of predicting battery voltage with forest [78], Gaussian process regression (GPR) [79], and multiple inputs [71] and forecasting parameters such as voltage, support vector machine (SVM) [80].

These algorithms are temperature, and SoC simultaneously [72]. However, LSTM commonly adopted partially because of their simplicity, which being a type of recurrent neural networks (RNNs) presented however is one of the reasons that they cannot handle the challenges in practical training scenarios, with poor training complex battery operating conditions. For relatively more stability and issues such as vanishing or exploding gradients. complex algorithms, the CNNs [81] capture local feature b) AE and VAE-based Approaches: GenAI can po- representation for time series forecasting and may overlook tentially address the above-mentioned challenges and the distant variable correlations, limiting the ability to capture reconstruction-based models have been studied.

AE as a basic remote topological structures. The long dependencies can reconstruction model is shown to be ineffective for generating be learned by LSTM [82], which however is limited with diverse and high-quality data. This is largely due to the extended sequences due to the inherent constraint of recurrent deterministic nature of the latent codes produced by the AE models, where signals must traverse both forward and backward. encoder.

A more suitable model is VAE, which adeptly learns Moreover, traditional ML models typically require extensive the probability distribution of multivariate time series (MVTS) data for model training; otherwise, the model could be over- to be robust against perturbations and noise [73]. In [51], a semi- fitted with reduced accuracy. Unfortunately, data is often scarce. supervised VAE-based anomaly detection model was proposed One idea is to use GenAI for data augmentation, and this idea for early anomaly detection in battery packs.

The model has been studied for SoC estimation, as detailed below. detects different anomalies such as irregular terminal voltage, b) GAN-based Approaches: Efforts to utilize GenAI for differences between all bricks, and abnormal temperature general-purpose data augmentation have yielded successes, e.g., fluctuations. The model, named GRU-VAE, consists of a gated convolutional recursive GAN (CR-GAN) [83], recurrent con-

ditional GAN (RC-GAN) [84], time-series GAN (TimeGAN) tasks. Generally, the common issues of the GAN-based models, [85], and global trend diffusion algorithms [86]. Among the such as training stability, also apply to TS-DCGAN, which GAN variants, CR-GAN and RC-GAN struggle to capture the can be further enhanced. temporal dynamics encompassed within the entirety of a time 3) State of Health (SoH) Estimation: EV owners are not series sometimes, owing to gradient vanishing or explosion that only interested in knowing the battery’s SoC but also SoH [91]. arose when processing long sequences with RNNs.

TimeGAN This allows them to plan effectively and schedule maintenance divides long time series into smaller segments, potentially or replacement as needed. It also offers vital insights into leading to the loss of crucial cross-segment information. The EV battery control strategies, protection mechanisms, and global trend diffusion algorithm lacks sufficient adaptability to sustainable development [92].

The SoH estimation methods can different scenarios because of its triangle distribution and fails be categorized into model-driven and data-driven approaches. to adequately represent the complexity of the original data. The former includes the electrochemical model [93] and Several GAN-based models are specially designed for SoC the equivalent circuit model [94]. Relatively, the latter has estimation.

The Wasserstein GAN (WGAN) is a promising advantages such as independence from prior knowledge of base model, which is shown to be robust by generating the battery mechanisms and avoidance of subjective intervention, underlying real data distribution, enhancing the generation with a focus on latent input-output relationships. quality of vanilla GAN, and accelerating convergence [87]. a) Traditional ML Approaches: SoH prediction has Hence, the authors in [65] developed a time-series Wasserstein been addressed by different traditional ML algorithms, e.g., GAN (TS-WGAN) based on WGAN for SoC estimation.

The artificial neural network (ANN) [95], SVM [96], automated new model consists of data pre-processing and a Wasserstein machine learning (AutoML) [97], GPR [98], CNN [99], RNN GAN with gradient penalty (WGAN-GP) architecture [87]. [100], LSTM [101], GRU [102], and Bayesian neural network Besides, the model is trained by the EV dataset [88] and LG (BNN) [103]. The above models are regarded as discriminative 18650HG2 Li-ion Battery dataset [89].

The former includes EV models, which focus on battery health indices including cycles information during charging/discharging such as timestamp, ve- and capacity. They map input parameters to output variables hicle speed, voltage, current, cell temperature, motor controller without prior sample knowledge, and adjust network weights input voltage and current, mileage, and SoC. The latter consists through training with a loss function.

However, it is challenging of the battery’s performance data during charging, discharge for them to well capture intrinsic characteristics to represent cycle measurements, and drive cycles. Note that TS-WGAN the battery’s operational dynamics accurately, and the challenge may suffer from convergence issues as it requires complex is often addressed by increasing the quality and quantity of index modifications due to the nature of GANs. the training dataset.

Another WGAN-based model is proposed in [66], named b) Diffusion-based Approach: Compared to conven- conditional LSTM Wasserstein GAN with gradient penalty tional discriminative ML models, GDMs, can capture the (C-LSTM-WGAN-GP). It is an LSTM-based conditional GAN distribution characteristics inherent in training data more accu- model and owns the capability to generate data that closely rately, thereby offering a more comprehensive understanding resembles actual battery data across various profiles.

The of the underlying problem. By leveraging generative diffusion training data is obtained from the experiments. Two Li-ion techniques, one could mitigate the risk of introducing significant rechargeable cells were set up for the experiments where one deviations to the overall distribution of training data, thereby is Li-ion with the material of Ni/Co/Mn ternary composites facilitating robust modeling that transcends merely isolated and another one is Li-ion with the material of lithium iron feature representations [104]. phosphate.

The usable battery information, such as terminal The authors in [60] introduced a diffusion-based model voltage, current, and temperature, was monitored and recorded namely, DDPM, to predict the SoH of lithium-ion batteries while SoC was estimated after the experiment was done. Future with both offline and online modeling. The dataset used in improvements can be architecture and training convergence their experiment is a lithium iron phosphate battery dataset optimization, as well as system integration with existing BMS. from TOYOTA Research Institute [105] and a nickel cobalt Extending from the above WGAN-based models, the manganese battery dataset from their laboratory [60].

The authors in [36] introduced a time-series deep convolutional datasets consist of specifications such as rated capacity, number GAN (TS-DCGAN) framework. The framework combines the of cells, charging/discharging current, maximum/minimum cut- time-frequency domain techniques and DCGAN [29] to train off voltage, and number of cycles. The prediction variables are the models for SoC estimation.

The models generate synthetic the battery’s capacity in the unit of Ah. The proposed DDPM datasets with high fidelity and diversity, effectively capturing outperforms other ML techniques in terms of several SoH the dependencies between multidimensional time series. The prediction error metrics, e.g., root mean square error (RMSE), training data is obtained from LG INR18650HG2 batteries [90], mean absolute error (MAE), and mean absolute percentage which are common EV batteries [36].

The dataset consists of error (MAPE). Taking RMSE as an example, DDPM can reduce battery voltage, current, cell temperature, SoC at different prediction errors of RNN by 57%, LSTM by 35%, GRU by temperatures, and driving cycles. Real data is complemented 10%, transformer [106] by 70%, and CNN-Transformer [107] with the generated synthetic data.

The experimental results by 52% [60]. In the future, DDPM’s effectiveness for SoH show that TS-DCGAN successfully reduces the discriminative may further be explored by comparing it with other GenAI score of CR-GAN by 53% and TimeGAN by 46% in producing methods such as VAE and GAN. reliable synthetic datasets for the subsequent SoC estimation Overall for layer 1, we provide a summary of the GenAI

TABLE II

S UMMARY OF G ENAI MODELS FOR I O EV IN LAYER 1 FOR BATTERIES . ●: G ENAI METHODS ; ✓: PROS OF THE METHODS ; ✗: CONS OF THE METHODS .

Application Reference GenAI Model Pros & Cons

● A GRU-VAE framework for battery anomaly detection. ✓ Learn probability distribution of multivariate time series data adeptly. [51] GRU-VAE ✓ Robust against perturbations and noise. ✗ Adaptability issues and frequent model updates for new data. Anomaly Detection ● A transformer-based method for battery fault diagnosis and failure prognosis. ✓ Learn battery’s nonlinear cell behaviors in a self-supervised data-driven manner. [25] BERTtery ✓ Competitive performance, e.g., above 95% in accuracy, precision, recall and F1 score. ✗ Lack generalization ability to various battery types and operational conditions. ✗ Need for an improved early warning predictions with reduced false positives and negatives. ● A GAN-based approach for SoC estimation of lithium-ion batteries. ✓ Robust and able to generate underlying data distribution. [65] TS-WGAN ✓ Enhanced generation quality of vanilla GAN and accelerated convergence. ✗ Convergence issue during training and requiring complex index modifications. ✗ Limitations for real-time SoC estimation due to its computational intensity. ● A GAN-based approach to generate synthetic data for SoC estimation.

SoC [36] TS-DCGAN ✓ Produce synthetic data of high fidelity and diversity. Estimation ✗ Common issues of GAN-based structure, e.g., stability of model training. ● A LSTM-based conditional GAN model for data generation and SoC estimation. ✓ Stable model training performance. C-LSTM- [66] ✓ Generate realistic battery data cross various profiles.

WGAN-GP ✗ Need for an improved model architecture and the speed of convergence. ✗ Lack of implementation on online BMS. ● A diffusion-based model for battery SoH estimation. SoH [60] DDPM ✓ Low prediction errors compared with RNN, LSTM, GRU, and transformer-based models. Estimation ✗ Need for a comparison with other GenAIs, e.g., GAN and VAE for the same application.

works in Table II. From the table, we can see that the Target Dataset (small) Source Dataset (large) Date, Year Date, Year VAE-based [51] and transformer-based [25] models can be Arrival time Arrival time utilized for battery anomaly detection. Moreover, GAN-based Plug-out time Required Energy Plug-out time Required Energy models, e.g., TS-DCGAN [36] and C-LSTM-WGAN-GP [66], are developed mainly for data augmentation to enhance the Data Augmentation Data Preprocessing via GAN accuracy of SoC estimation.

And TS-WGAN [65] considers both data augmentation and SoC estimation. Furthermore, the Augmented Train Source Train Source recent advancements in generative diffusion models have been Target Dateset DNN DNN applied in DDPM [60] for SoH estimation. Predicted Plug-out Predicted Design Target DNN Time of EV Owners Required Energy (Plug-out Time) B.

Layer 2: Individual Electric Vehicle (EV) Transfer Learning Process Transfer Learning Process

Fine-tuning EV is an integrated system that combines key components Target DNN Source DNN Target DNN Source DNN Target DNN like batteries in layer 1. Instead of focusing on one component Predicted Plug- as the research works for layer 1, the research in the EV out time of Design Target DNN (Required Energy) target dataset layer emphasizes the integration and functionality of the EV system. GenAI has been studied for the EV layer for two Predicted Fine-tuning Required Energy Target DNN aspects including charging [17], [108] and routing [26], and we present technical details below.

Fig. 4. The flowchart of proposed framework to address the cold-start 1) EV Charging Behaviors and Loads: The research on EV forecasting problem in predicting the EV charging behaviors such as plug- charging is important for optimizing energy usage, balancing out hour and required energy for newly committed EVs [17]. grid demand, and improving charging efficiency. The growing adoption of EVs makes data-driven approaches ideal for related research with a growing amount of EV data generated.

The management for EVs, e.g., RNN [110], CNN [111], CNN- approaches are typically designed for optimizing the charging GRU [112], and LSTM with RL [12]. A common issue of parameters and analyzing parameter correlations, user travel the above algorithms is the demand for extensive training patterns, and vehicle trajectories [109]. data (to avoid overfitting and underfitting) [113]. Such data a) Traditional ML-based Approaches: Several ML demand cannot be satisfied by many EV service providers, algorithms have been applied in load forecasting and energy facing realistic constraints, e.g., only 365 charging samples

per year. This is especially true at the beginning stages of data tasks, e.g., optimal scheduling with HEDGE generated profiles. collection, causing the famous “cold-start forecast problem” c) Hybrid Approach: Instead of using GAN alone, [114]. Therefore, researchers are looking for methods to hybrid solutions have been investigated for synthetic data generate extensive datasets with relatively small-scale data generation also.

In [108], a VAE-GAN model is developed to collected from real EVs, and GenAI is a promising method. generate synthetic time-series energy profiles such as EV load b) GAN-based Approaches: The above-mentioned cold- profiles in smart homes. The synthetic data is subsequently start problem has been addressed in [17]. The authors developed utilized in the Q-learning-based home energy management a transfer learning-based framework using a deep generative system (EMS) to maximize long-term profit through optimal model, GAN, to address the problem of predicting EV load scheduling.

The model is trained with the iHomeLab PART charging behaviors. Fig. 4 shows the flowchart of the proposed dataset [123], which includes power consumption profiles of framework which integrates GAN and DNN for the forecasting five residences. The study compared the VAE-GAN with a tasks.

As seen from the figure, two source DNNs are trained Vanilla GAN and a GMM, employing the Kullback-Leibler on residential EV owner data to predict the plug-out time of (KL) divergence to assess the distance between real and EV owners and the required energy, respectively. The transfer synthetic data distributions. The results reveal that VAE-GAN learning then adapts the knowledge of the tested EV to the can achieve 18% and 33% performance improvement over target DNN (plug-out time) for forecasting new EV plug-out GAN and GMM, respectively, in generating EV load data hours.

The same approach is applied to the target DNN for [108]. The improvement shows that the model is able to learn required energy. Also, GAN models are used to augment the various smart home data distributions (e.g., electric load, PV target dataset.

The target DNNs, with the weights from the generation, and EV charging load), and generate realistic data source EV model and the GAN-generated data, are fine-tuned samples without prior analysis before training. Nonetheless, the to predict the charging behaviors of target EVs. model relies on the quality and diversity of training data with Besides, the research is based on a public dataset from EA inconsistent scalability and adaptability.

Future research may Technology, which specializes in providing asset management incorporate diverse datasets to enhance the model’s robustness solutions for the owners and operators of electrical assets [115]. and improve scalability across broader smart grid applications. The available parameters of residential charging events include 2) EV Routing: The future of urban delivery is likely to be dates, arrival hours, plug-out hours, and required energy.

Finally, driven by autonomous green vehicles [124]. The trend under- the derived models can achieve significant performance gains scores the significance of efficient route planning, addressed over support vector regression (SVR) [116] by 12%, decision as the electric vehicle routing problem (EVRP). In EVRP, an tree regression [117] by 57%, k-nearest neighbor regressor EV starts from a designated depot with a partial/full charge (KNNR) by 60%, DNN [112] by 31%, and a GAN-DNN based to serve customers with time restrictions.

Each EV can stop approach [16] by 10% [17]. For future work, the performance of at charging stations or return to the depot to recharge. The proposed method could be improved by applying the clustering goal of EVRP is to find cost-effective routes for the EV fleet algorithms for efficient transfer and multi-source datasets for subject to battery constraints. model generalizability. a) RL-based Approaches: Recently, researchers have In [118], a convolution conditional GANs with Wasserstein applied supervised learning and RL to address the vehicle distance as network objective function (CW-GAN) model was routing problem (VRP) amidst the growth of ML, e.g., a developed for generating EV charging behavior parameters such pointer network [125].

A significant challenge is to obtain as arrival time, departure time, and SoC. The study is based sufficient labels for producing optimal solutions in large-scale on a small dataset of private EVs and charging piles in three problems like EVRP. RL is a label-free approach so it can functional zones (i.e., office, business, and residential areas) be a viable option for addressing large-scale problems [126]. [118].

The input of the model includes noise and conditional Deep reinforcement learning (DRL), as a type of RL, has been labels, and the output generates different parameters of EV applied to solve the VRP, often utilizing the encoder-decoder charging behavior. Note that the quality of conditional labels architecture in neural network design [127], and achieved good has a significant impact on the overall performance of CW- performance.

However, related research works often focus GAN, so accurate label selection is crucial. on basic routing issues, and overlook the complexities of Besides charging behavior, GAN has also been studied EVRP with energy and charging constraints, which are unique for charging load. In [119], a GAN-based home electricity compared to traditional VRP. data generator (HEDGE) tool is introduced to semi-randomly b) Transformer-based Approach: Transformer architec- generate synthetic daily profiles of EV and household loads ture with an attention mechanism has been proven to be effec- as well as PV generation.

The generated residential energy tive in improving computational efficiency and solution quality data spanning multiple days exhibits consistency in terms of in solving VRP [128]. Hence, the authors in [26] proposed a magnitude and behavioral clusters. Several profile datasets are Transformer-based DRL method for energy minimization of used for model training.

Household load and solar generation EVRP. Fig. 5 shows the proposed framework where the policy are available in TC1a [120] and TC5 [121], respectively, from network of DRL is modeled by the Transformer’s encoder- the customer-led network revolution (CLNR) project. The EV decoder structure.

The features of EVRP are captured by the loads are estimated based on the general population’s travel feature embedding module and the policy gradient method patterns dataset from the UK’s National Travel Survey [122]. is employed for policy training. In the context of an EVRP In the future, HEDGE will be valuable for subsequent research instance, the graph information including the vehicle and the

AV-related aspects such as trustworthiness and navigation. Policy Transformer Overall, the GenAI applications for IoEV in layer 2 are State Network -Based summarized in Table III. We find that the GAN-based models, Step 1 Update Vertices GAN-DNN [17] and CW-GAN [118], can be utilized to Encoder State generate EV charging behaviors.

For residential applications, Feature the GAN-based models [108], [119] are employed for data Vehicle State Embedding augmentations, which is important for other applications, e.g., EVRP Module Environment optimal scheduling of an EV at home [108]. Researchers also use the Transformer model [26] to solve the EVRP, which is Step 2 Strategy Decoder different from traditional VRP with new EV constraints.

C. Layer 3: Smart Grid with EV

EVRP The proliferation of EVs leads to increased stochastic power Environment demands from the grid, accelerating grid asset deterioration Update Policy Solution Baseline and complicating power system operations. The investigation Network by Policy Network of EV charging load profiles is crucial for understanding future Step t ··· grid states to enable large-scale transportation electrification. However, the issues persist regarding the quantity and quality Step T of data [14], [16], and the challenges involved in predicting Choose EV charging loads [15], [27], [38], [136], [137] as well as the End of Actions Loss Episode Greedly urge of understanding user experience [24]. 1) Data Quantity and Quality: The digitalization and EVRP Environment widespread deployment of charging infrastructure offers a great Policy Gradient Method opportunity to gather real EV charging data, yet it is hampered by equipment failures, data collection errors, and intentional Fig. 5.

Framework of DRL with Transformer for EV routing problem [26]. damage, resulting in missing values and outliers [138]. Given insufficient data accumulation in newly built charging facilities, the ML models can be biased with the flawed datasets [139], states of vertices undergo encoding by an encoder. This process which are not necessary to be small-scale.

Such bias and facilitates the step-by-step construction of EV routes by the inaccuracy pose challenges to the scheduling and optimization decoder, leveraging inputs from both the encoder and the feature of the grid, whether centralized or distributed. As such, one embedding module. Subsequently, updates to the parameters of important usage of GenAI is to enhance the EV and grid the policy network are made based on reward values derived datasets to improve the system performance such as load from the policy network and the baseline network. forecasting and balancing.

Several GenAI algorithms, such The EVRP instances are created according to the procedure as VAE [14] and GAN [16], [38], [140], have been explored outlined in [129]. The locations of customers and charging and we present the technical details of them below. stations were chosen uniformly at random from a square a) VAE-based Approach: VAE’s adoption is mainly kilometer area. Finally, the proposed transformer-based DRL for generating stochastic scenarios for EV load profiles.

In method is compared with the exact algorithm [130], improved [14], a VAE model is designed for such usage to capture ant colony algorithm (ACO) [131], adaptive large neighborhood the time-varying and dynamic nature of EV loads effectively. search (ALNS) algorithm [132], and DRL with general attention The paper considers five different EV load profiles, for fully model (AM) [128] under various scenarios, e.g., different battery-based and hybrid-based EVs with and without demand number of EVs and charging stations.

Specifically for a case responses [14]. The profiles are measured at 10-minute intervals, study with 100 EVs, the proposed method can reduce the resulting in 144 data points for each profile per day. The model energy consumption of EVs by 1% compared to AM, and 10% uses the historical profiles as input, and accordingly generates to ACO [26].

Future work includes enhancing the model’s new profiles that encapsulate the critical characteristics of efficiency, testing with real-world data, and expanding the the profiles. For future work, the proposed method holds the model to accommodate more complex scenarios, e.g., multiple potential to be integrated with load forecasting models [27] as vehicle types and different environmental constraints. Other a data augmentation tool to address the data scarcity issue.

GenAIs such as GAN and GDM can also be considered for b) GAN-based Approaches: Similar to VAE, GAN has navigation and route optimization application [133]. also been applied for data augmentation and furthermore, load At the end of such EV layer, we would like to mention forecasting. In [16], GAN is used to generate EV charging load Autonomous Vehicles (AVs), which often occur with EV data first, with which, load forecasting is performed.

The data together as new concepts of vehicular technology. However, an augmentation model is called GRU-GAN. The model uses the AV is not necessarily an EV and it can be driven by different transactional data from various EV charging stations within a power sources including but not limited to electricity.

Our 35 kV power distribution zone spanning from July 1 to August focus in this paper is EV and some interesting discussions 31 in 2019 [16]. Due to different realistic restrictions, the time- of utilizing GenAI in AVs are available in [134], [135] about series data is incomplete. As such, the data is pre-processed to

TABLE III

S UMMARY OF G ENAI FOR I O EV IN LAYER 2. ●: G ENAI METHODS ; ✓: PROS OF THE METHODS ; ✗: CONS OF THE METHODS .

Applications Reference Techniques Pros & Cons

● A framework to predict EV charging behaviors, e.g., plug-out hour and required energy. EV Charging [17] GAN-DNN ✓ Address the cold-start forecasting problem when limited training data is available. Behavior ✗ Need clustering algorithms and multi-source datasets for future work.

Data Aug- ● A GAN-based model to generate charging behavior data for EVs. mentation [118] CW-GAN ✓ Learn from small samples, expanding dataset while preserving original probability distribution. ✗ Dependency on quality of conditional labels during the training. Smart Home VAE-GAN ● A data generation scheme using a VAE-GAN combined with Q-learning-based home EMS. Data Aug- [108] with ✓ Learn various data distributions in a smart home and generate realistic samples. mentation Q-learning ✗ Reliance on quality and diversity of training data.

Residential ● A GAN-based tool for semi-randomly generated data for EV load, PV, and household demand. EV Load [119] GAN ✓ Generate profiles that keep both magnitude of profile and behavioral consistency over time. Generation ✗ Need other countries’ datasets to extend the capability of the HEDGE tool. ● A Transformer-based DRL method for solving electric vehicle routing problem (EVRP). ✓ Consider EV’s energy and charging constraints.

Transformer- EVRP [26] ✓ Reduced energy consumption of EV compared with exact algorithm, ACO, ALNS, and AM. based DRL ✗ Lack model testing with real-world data. ✗ Need to consider data diversity, e.g., various EV types.

simultaneously identify and handle missing values and outliers correlation among spatial regions. using GRU-GAN, and this process is commonly referred to To address the aforementioned challenges, the Transformer as data imputation. Finally, the generated high-quality data is models [15], [27] and the GAN models [38], [136], [137] used for training the Mogrifier LSTM network for short-term have been introduced recently.

The former utilizes attention EV load forecasting. mechanisms to capture the long-range dependencies and The comparison results show that GRU-GAN outperforms intricate patterns of the time-series data. This allows the model conventional imputation techniques. Compared to mean impu- to focus on relevant parts of the input sequence and facilitate tation [141], GRU-GAN can reduce the errors by 15%.

GRU- more accurate and robust load forecasting results. The latter GAN’s performance improvement is even more significant helps to generate highly realistic EV charging scenarios by when comparing with piecewise linear [142] and k-nearest learning the underlying data distribution over a latent space neighbor (KNN) [143] imputation, achieving 43% and 19% with conditions. A wide range of plausible scenarios generated lower errors, respectively.

Such enhanced imputation accuracy by GAN can provide a comprehensive view of potential future is attributed to GRU-GAN’s ability to capture intricate high- loads. We present the technical details as below. level data representations. The potential improvement of the b) Transformer-based Approach: The transformer- solution is mainly in the forecasting stage.

The forecasting of based model has been widely used for time-series forecasting certain intervals is especially challenging, e.g., peak periods. because it addresses the limitations of LSTMs and effec- The GenAI models, e.g., transformer-based [15], [27], may tively captures long-term dependencies [15]. The integration serve as viable substitutes for Mogrifier LSTM. of a transformer-based architecture with the probabilistic 2) EV Charging Load Prediction: Understanding the forecasting technique in the temporal latent auto-encoder pattern of EV load is critical for the grid scheduling.

Con- exhibited significantly enhanced efficacy in the context of ventionally, the load prediction is formulated as a time series time series forecasting endeavors [149]. Nevertheless, the forecasting problem, which can be handled by different ML vanilla transformer model exhibits significant time and memory algorithms such as the GenAI-based ones. overheads due to the quadratic computation complexity of self- a) Traditional ML-based Approaches: There exist attention, and imposes constraints on the maximum allowable several load forecasting approaches driven by traditional ML, input sequence length as a result of the accumulation of e.g., SVR [144], ANN [145], RNN [146], and LSTM [147]. encoder and decoder layers [150].

The Informer [150] tackles These algorithms have demonstrated the potential of using these challenges by substituting the conventional self-attention ML for load forecasting. However, they either fell short in computation in the standard transformer with a ProbSparse capturing the nonlinear characteristics inherent in EV load self-attention framework [150], while also introducing the self- series or faced challenges in modeling long-term dependencies attention distilling mechanism.

However, it is limited to point that are common in real-world forecasting applications [107]. In forecasting which refers to the prediction of a single/specific the application layer, these algorithms focus on the load patterns value for a future variable or event [151]. in time scale where spatial load distribution is overlooked. A The authors in [27] proposed a Probformer model to address recent work [148] developed a spatial–temporal forecasting the challenge of capturing long-term dependencies within method for load forecasting, but the method cannot model the charging load sequences and to facilitate the generation of prob-

abilistic load forecasts. The model with multi-head ProbSparse 3) Consumer Data and Sentiment Analysis: The trans- self-attention [150] was an adaptation of the Transformer-based portation sector significantly contributes to global greenhouse Informer framework [150]. Subsequently, the MetaProbformer gas (GHG) emissions [155] and impacts public health [156]. [27] was developed by integrating the Probformer with a meta- Government policies increasingly favor EVs to reduce GHG learning algorithm, Reptile [152], which tackled the issue emissions.

However, the analysts underutilized consumer data, of charging stations with scarce historical data. Four datasets particularly unstructured EV data, in decisions about charging comprising real-world EV charging loads, sourced from various infrastructure. Previous research employing sentiment analysis charging stations over distinct time periods, were utilized in the suggested prevalent negative user experiences in EV charging experiments.

Each dataset details the start and end times, along station reviews, yet lacked specific causal extraction [157]. with the total energy consumed (in kWh) for each charging Thus, the multi-label topic classification is crucial for under- event. While MetaProbformer is capable in load forecasting standing user interaction behaviors in electric mobility. Hence, with limited historical data, it faces difficulties for new charging the authors in [24] utilized the Transformer-based models, Bidi- stations where data collection takes time.

Hence, an exploration rectional encoder representations from Transformers (BERT) of methods for enhancing the model generalizability within a [158] and XLNet [159], for the multi-label topic classification few-shot or zero-shot framework is necessary. in the domain of EV charging reviews. The method proposed c) GAN-based Approach: Recently, the GAN-based herein aimed to expedite the evaluation of research through techniques for charging scenario generation have begun to automated means, utilizing extensive consumer data to assess emerge [38], [136], [137].

For example, the authors in [136] performance and analyze regional policies. The study utilized proposed WGAN-GP to tackle spatial–temporal uncertainty data derived from 12,720 charging station locations across the in EV charging load analysis. This approach explored load United States, comprising 127,257 English-language consumer dynamics and generated scenarios without uniform probability reviews written by 29,532 EV drivers over a four-year period assumptions across charging stations.

The undisclosed power from 2011 to 2015 [160]. An important future research is to grid structure limited data access to the EV charging load. To enhance model interpretability through methods such as the model the impact of these stations on the distribution network, use of rationales [161], influence functions [162], and sequence the data from 32 charging stations in the Zhejiang region tagging approaches [163] to well understand consumers. were allocated to nodes in the IEEE 33-node distribution The GenAI applications for IoEV in layer 3 are summarized network system [136].

The forecasting target was the EV in Table IV. We observed that the GAN models were often charging load at each node in the network. Nevertheless, the used for data augmentation [38], [140], handling missing values comparative analysis of the WGAN-GP against contemporary and outliers of the input data [16], and generating EV charging GenAI approaches in the context of EV charging scenario sessions/scenarios [136], [137].

Besides GAN, VAE can be generation remains lacking. utilized for generating stochastic scenarios for EV load profiles. A Copula generative adversarial network (CopulaGAN) The Transformer-based models were usually utilized for the model combining Copula transformation with GANs was multi-label topic classification in the domain of EV charging developed in [137]. Slightly different from the traditional reviews [24], and EV load forecasting [15], [27]. load forecasting approaches focusing solely on EV charging load curves, the CopulaGAN captured uncertainties in EV charging sessions, including energy delivered, arrival, and D.

Layer 4: Security departure times. Subsequently, it was used for day-ahead Security is a critical aspect that we cannot bypass for many optimal EV scheduling. The charging session dataset used cyber-physical systems.

In the following parts, we introduce in this study was obtained from the Caltech parking lots [153]. the background of security research for electric mobility and Each site’s database records the EV connection time, charging then present related GenAI research works. completion time, energy consumption, and vehicle departure 1) Background: Security’s impact on electric mobility time from the parking lot. The power market dataset including encompasses several aspects.

The first one is due to the ever- the load forecast, generation forecast, day-ahead electricity increasing adoption of the ML models. For example, there has market price, and balancing energy market, was collected been considerable utilization of DRL algorithms within the from the German electricity market [154]. The simulation of context of EV charging schedules, aiming at acquiring optimal EV charging sessions and the prediction of day-ahead market charging strategies for users.

Such algorithms include soft actor- prices were carried out for 48 hours. The outcomes of their critic (SAC) [164], deep deterministic policy gradient (DDPG) investigation demonstrated a high degree of correspondence [165], safe deep reinforcement learning (SDRL) [166], multi- between the generated EV charging session data and the actual agent deep reinforcement learning (MADRL) [167], proximal dataset, with an approximate match rate exceeding 90% [137]. policy optimization (PPO) [168], etc.

Same as many ML This validation underscored the effectiveness of CopulaGAN algorithms, DRL suffers from potential adversarial attacks in accurately capturing the inherent structure and distributional [169]. The adversarial examples involve maliciously altered attributes of the data. However, the proposed optimal scheduling inputs to deceive the ML models, causing erroneous outputs of EVs charging faces limitations due to data inadequacy and [170].

Some methods are shown to be effective in generating assumptions about SoC and charging rates. The comparative adversarial data, e.g., fast gradient sign method (FGSM) [171], analysis of CopulaGAN and other GenAI methods (e.g., GAN- basic iterative method (BIM) [172], and DeepFool [173]. The DNN [17] and DiffCharge [61]) shall be investigated. ramifications of such attacks could be extensive, ranging from

TABLE IV

S UMMARY OF G ENAI FOR I O EV IN LAYER 3. ●: G ENAI METHODS ; ✓: PROS OF THE METHODS ; ✗: CONS OF THE METHODS .

● A framework for generating and enhancing EV load profiles. ✓ Ensure consistency in power consumption between generated and original profiles over time. [14] VAE EV Charging ✓ Capture temporal correlations, probability distributions, and volatility of original load profiles. Load Profile ✗ Challenge in balancing reconstruction loss and KL divergence during training. Data Aug- ● A load data generation model for missing values and outliers of the input data. mentation GRU-GAN ✓ Generate high-quality data closely resembling real data. [16] with Mogrifier ✓ Outperform traditional mean [141], piecewise linear [142], and KNN [143] imputation methods LSTM ✗ Sub-optimal forecasting performance during peak and plateau periods.

Analysis of ● Transformer-based models for analyzing EV charging reviews. EV [24] BERT, XLNet ✓ Outperform traditional LSTM and CNN models in terms of accuracy and F1 scores. Consumer ✗ Lack of interpretability. ● A Transformer-based model for EV charging load forecasting. ✓ Perform well in point and probabilistic forecasting for the load at EV charging station.

EV Load ✓ Competitive performance in probabilistic forecasting for both short-term and long-term tasks. [27] MetaProbformer Forecasting ✓ Adaptable to seen and unseen scenarios. ✗ Need an extension to multivariate forecasting for future work. ✗ Need to improve model’s generalization capabilities within a few-shot/zero-shot framework. ● A GAN-based model for generating EV charging scenarios at charging stations. [136] WGAN-GP ✓ High degree of spatial correlation similarity in terms of SSIM and FSIM indexes.

Charging ✗ Lack of comparative analysis for different GenAI methods in EV charging scenario generation. Scenarios ● A GAN-based model for capturing and modelling the uncertainties in EV charging sessions. Generation ✓ Generate EV charging sessions data which closely align with the real dataset. [137] CopulaGAN ✗ Real conditions not fully captured with simplified assumptions of SoC and charging rate. ✗ Need a comparison study with other GenAI approaches for scenarios generation.

increasing charging expenses and fluctuations in grid loads to mechanisms, susceptibility to multiple attack vectors, and jeopardizing the stability of power grids [174]. the absence of encryption technologies [183]. The security vulnerabilities have been studied with several methods such as Besides adversarial attacks, security is also a concern in local outlier factor (LOF), one-class support vector machines power grids with EV penetration.

Specifically, the integration of (OCSVM), and principal component analysis (PCA) [184]. power systems with electrified transportation networks poses Nevertheless, these methods often are sensitive to noise, new challenges concerning the reliability and resilience of incapable to handle high-dimensional data, and fail to handle charging infrastructure [175]. According to [176], there is an the intricate dynamics of systems. increase in the occurrence of power system attacks targeting customer satisfaction levels in charging services.

Deep learning 2) Adversarial Attacks: To investigate adversarial attacks algorithms such as deep belief network (DBN) [177] and against DRL in the EV charging process, a GAN-based feed-forward neural network (FNN) [178] have been used approach namely RL-AdvGAN was introduced in [39]. RL- for detecting cyber-physical attacks in power systems because AdvGAN could support the adversary to leverage the stolen of their advanced feature extraction capabilities.

The algorithms data to engage in behavior cloning, thereby constructing an achieve high detection rates, yet they overlook the extraction adversarial policy network to mimic the user’s policy. In this of the important spatial relationships inherent in the data, as research, the EV charging environment was set up using real- they disregard the topological grid attributes [179]. world data from California independent system operator (ISO) The connection between EVs and the grid is largely [185].

The dataset spanned a duration of approximately 2 controlled by the EV’s supply equipment, that is responsible years with data recorded hourly, consisting of the electricity for managing and maintaining the charging operations. It prices and the grid load information. The paper assumed also facilitates communication among cloud services, payment that the commuting behavior of EV users follows the normal providers, EVs, battery management systems, and other relevant distributions with parameters of arrival time, departure time, entities to enable efficient and intelligent charging [180]. and battery SoC.

Four types of DRL algorithms, i.e., deep However, the connection is often associated with potential Q-networks (DQN) [186], DDPG [187], PPO [188], and SAC vulnerabilities such as DoS attacks, FDIAs, and spoofing [189] were tested under the adversarial attacks generated by [181]. A popular connection standard is CAN, which is a FGSM [190] and RL-AdvGAN [39]. According to the results, bus protocol for communication within vehicles.

CAN has FGSM attack is effective, and the algorithms, i.e., DQN, DDPG, many advantages such as reduced wiring expenses, minimal PPO, and SAC, only managed to maintain a normal SoC weight, and simplified design [182]. However, CAN has range of the battery in 6%, 0%, 28%, and 8%, of the time, security vulnerabilities also, including inadequate authentication respectively [39]. The proposed RL-AdvGAN attack [39] was

even more effective, where all the tested DRL algorithms failed and decision-making in dynamic operational contexts of power to maintain the normal range of SoC all the time during the and transportation networks. attack. As such the proposed RL-AdvGAN posed a significant Based on the above introduced detection model, the re- risk to the security and stability of the battery system. Future searchers in [47] developed an interpretable anomaly detection work includes further exploring the threat and damage of system, referred to as RX-ADS.

The system was designed to GenAI methods for adversarial attacks with different network identify intrusions within the CAN protocol communications of architectures and developing effective attack detection methods. the EVs with active charging connections. Besides, the ResNet In [191], the authors considered both the generation and Autoencoder was employed to learn normal behavior from detection of adversarial attacks against vehicle-to-microgrid data and detect anomalies based on reconstruction errors.

The (V2M) systems and proposed GAN-based models. The re- performance of the system was investigated with two publicly search modeled a system where the adversaries sought to available datasets of EV CAN protocol, including offset ratio manipulate the ML classifier at the network edge, causing and time interval based intrusion detection system (OTIDS) it to incorrectly classify the energy requests received from [195] and Car-Hacking [196].

The results showed that the microgrid users, e.g. EVs’ charging/discharging requests. The system outperformed a GAN-based intrusion detection system FGSM and conditional generative adversarial network (CGAN) [197] by 4% under DoS attacks.

Nevertheless, the proposed models were introduced to serve as attackers in generating method requires a substantial dataset that accurately reflects the adversarial instances. On the adversarial detection side, a system’s typical normal behavior. If new behaviors arise that GAN-based adversarial training framework was introduced deviate from the established norms, the model shall be updated. to create adversarial training instances.

The instances are used Otherwise, such deviations might be incorrectly classified as to train SVM classifiers to detect these adversarial attacks. The anomalies, leading to an increase in false positives. simulation was set up using the iHomeLab RAPT dataset [192] The same research group of [47] developed a ResNet AE- consisting of electrical power consumption and generation of based approach for unsupervised physical anomaly detection 5 Swiss households, and the supplementary dataset with power with high resilience in [198], specifically in EV charging generation from the batteries and wind farm from their previous stations without labeled data.

The experiments were conducted work [193]. The results indicate that their proposed method using data under normal and physical attack scenarios from outperforms the density-based spatial clustering of applications the Idaho National Laboratory’s EV charging station system with noise (DBSCAN) algorithm, leading to an improvement testbed. The proposed ResNet Autoencoder based approach was in the adversarial detection rate ranging from 13.2% to 25.6% compared with two benchmark algorithms: LOF and OCSVM. [191].

For future work, the investigation of the impact of limited The results in terms of F1 score show that the proposed method resources on the adversarial detection rate is necessary since is able to improve the detection performances of LOF algorithm the classifier is designed to be deployed at the network edge. by around 20.1% and OCSVM algorithm by about 3.7% [198]. 3) Detection of Attacks: The authors in [45] extended For further work, enhancing the proposed anomaly detection the study of potential attacks [39] and developed a graph framework can be achieved by integrating cyber security related autoencoder (GAE) model for detecting the FDIAs in power scenarios pertinent to EV charging systems. systems.

The model leverages the correlations between power In addition to detecting FDIAs [45], [199], the authors system data (e.g., active and reactive power measurements) in [46] proposed a ResNet AE-based anomaly detection and transportation data (e.g., hourly traffic volume) to enhance framework which consists of cyber anomaly detection (Cy- charging satisfaction. In the event of FDIAs, the malicious ADS) and physical anomaly detection (Phy-ADS) for the entities can manipulate power measurements to simulate cyber and physical data streams, respectively.

The former has additive, deductive, and camouflage attacks. The level of the capability to monitor and analyze packet data in real- satisfaction for customers can diminish due to insufficient time for the detection of cyber anomalous behaviors within power supply at the charging stations, e.g., blocked charging EV charging stations. Meanwhile, the latter is designed to requests or extended charging times.

The proposed method capture and analyze real-time physical sensor data (e.g., voltage, was tested via simulation of the Texas power grid consisting current, power, and thermal measurements) to identify physical of 2,000 buses and 360 active charging stations. The load anomalous behavior. Their results show that the proposed Cy- data was sourced from electric reliability council of Texas ADS for detection of cyber attack and Phy-ADS for physical (ERCOT) which manages the distribution of electric power to anomaly detection outperforms the LSTM AE method by about more than 27 million customers in Texas [194].

Compared to 18% and 15% respectively, VAE approach by around 3% and the benchmark graph CNN model [179], the proposed GAE 5% respectively [46]. However, the proposed method is subject model improved the detection accuracy by about 15% in various to the quality of the training data and model retraining is FDIAs scenarios on vulnerable nodes when 30% of data was needed if system behaviors change over time. under attack.

The results show that the proposed GAE model The GenAI research for EV and IoEV security in layer 4 is outperforms the state-of-the-art detector in various attacks. For summarized in Table V. As seen from the table, AE-based and example, compared to the benchmark graph CNN model [179], GAN-based models are commonly found in recent research the proposed GAE model can improve the detection accuracy for security.

Among them, GAN-based models are mainly by around 14.4% to 14.8% in the various FDIAs scenarios used for generating and identifying adversarial attacks, e.g., e.g, additive attacks, deductive attacks, and combined attacks RL-AdvGAN [39] and CGAN [191]. Whereas the AE-based [45]. Future research may focus on real-time model updates models [45]–[47], [198] are good at detecting FDIAs, DoS,

TABLE V

S UMMARY OF G ENAI FOR I O EV IN LAYER 4. ●: G ENAI METHODS ; ✓: PROS OF THE METHODS ; ✗: CONS OF THE METHODS .

● A GAN-based model for generating adversarial attacks against DRL algorithms. ✓ Effectively reduce the performance of common DRL algorithms for optimal EV charging. [39] RL-AdvGAN ✓ Higher security threat than FGSM [171]. Adversarial ✗ Need to improve training stability with model refinement. Attacks ✗ Need to propose methods to prevent adversarial attacks. ● A GAN-based detection framework for identifying the adversarial attacks.

GAN and [191] ✓ Higher adversarial detection rate compared with traditional DBSCAN algorithm [191]. CGAN ✗ Incompatible with network edge services due to resource constraints. False Data ● An AE-based detection scheme for identifying FDIAs.

Graph Injection [45] ✓ Performance improvement of 15-25% compared with SVM, FNN, CNN, and LSTM [45]. Autoencoder Attacks ✗ Need to update and deploy the offline-trained model to real-time applications for future work. ● An AE-based anomaly detection framework for identifying both cyber and physical attacks. Cyber and ResNet ✓ Capable of detecting both simple and complex cyber-physical attack scenarios.

Physical [46] Autoencoder ✓ Low training and inference time, making it suitable for real-time applications. Attacks (AE) ✗ Sensitive to the setting of threshold values. ● An AE-based method for detecting intrusions in CAN communication protocol for EV charging. DoS and ResNet ✓ Competitive results compared with the GAN-based approach [197].

Fuzzy [47] Autoencoder ✗ Require a large amount of data that reflects the normal behavior of the system. Attacks (AE) ✗ Need model update and retraining for new behaviors in the future. ResNet ● An AE-based approach for physical anomaly detection.

Physical [198] Autoencoder ✓ Outperform two benchmark algorithms including LOF and OCSVM [198]. Attacks (AE) ✗ Not capable of detecting cyber attacks.

fuzzy, and physical attacks, and serve as viable alternatives to curves. The daily charging load profile could be aggregated and traditional methods such as LOF and OCSVM. extracted from the ACN-Data dataset. Then, it was integrated with arrival/departure time and actual scheduled energy for E.

Studies Across Multiple Layers training the model to generate the EV load profile at the station. Besides studies on individual layer, some studies were car- The proposed method was compared to the baseline models: ried out on multiple IoEV layers, e.g., [40], [61]. Understanding GMM [203], VAEGAN [204], and TimeGAN [205].

Their user behavior [200], coupling characteristics [201] among user results showed that the proposed DiffCharge could generate behavior, road networks [202], and EVs are crucial for accurate realistic charging curves and it outperformed the baseline demand prediction, but it remains challenging due to various models in terms of marginal score, discriminative score, and factors such as time and SoC. Nevertheless, research on precise tail score.

For example, the marginal score of DiffCharge was mathematical models for charging and discharging strategies improved by 91% from GMM, by 35% from VAEGAN, and by is lacking due to the complexity of influencing factors. 32% from TimeGAN [61]. However, the control capability of a) Generation of EV Charging Scenarios: In [61], a the DiffCharge is restricted, thereby limiting the model’s ability diffusion model namely DiffCharge was developed to generate to generate tailored charging scenarios for varying conditions EV charging scenarios.

The generated scenarios could be such as initial SoC, types of batteries, and station congestion. divided into battery-level (e.g., charging current in Ampere) and b) Generation of Regional Electric Vehicle (EV) Charg- station-level (e.g., charging load in kW). The traditional ML ing Demand: Similar to [61], [40] explored the EV charging e.g., GMM could be utilized to estimate the daily EV charging demand from perspectives encompassing both battery and load profiles, but it faced a challenge in accurately capturing station levels.

However, a notable distinction of [40] lies in temporal dynamics across various time-series EV charging data. the emphasis placed on the spatial–temporal distribution of EV On the other hand, DiffCharge as one of the diffusion-based charging demand in the region. The authors in [40] addressed approaches is capable of deriving the challenging uncertainties the challenges of predicting spatial–temporal EV charging associated with charging and producing a range of charging load demand at both battery and station levels by proposing a deep profiles characterized by realistic and unique temporal features. learning framework that consists of methodologies such as On top of DDPM [52], the DiffCharge framework consists GAIL, PPO, and XGBoost.

The paper classified strategies of LSTM, broadcast, multi-head self-attention, and 1D-CNN. concerning user charging and discharging into three categories: DiffCharge was trained by using the ACN-Data [153] dataset driving policies, travel target mileage policies, and charging which comprises real-world charging data of individual EVs in duration selection policies. Following this categorization, the California. The data attributes including connection time, done research leveraged GAIL to obtain insights into these delineated charging time, kWh delivered, and charging current in Ampere policies i.e., GAIL was used as a strategy learning model.

Then, were considered for training the model to generate EV charging employing the PPO, the strategy learning model underwent

TABLE VI

S UMMARY OF G ENAI FOR I O EV ACROSS MULTIPLE LAYERS . ●: G ENAI METHODS ; ✓: PROS OF THE METHODS ; ✗: CONS OF THE METHODS .

● A diffusion-based model for generating EV charging scenarios for both battery-level (Layer 1) Scenarios and station-level (Layer 3). [61] DiffCharge Generation ✓ Outperform GMM [203], VAEGAN [204], and TimeGAN [205] in charging scenarios generation. ✗ Limited control over diverse conditions, e.g., initial SoC, battery types, and station congestion. SoC and GAIL, PPO, ● A GAN-based model, GAIL, as a strategy learning model assisting DRL algorithm.

Load [40] and XGBoost ✓ Good SoC prediction with low MAE and RMSE values. Forecasting [40] ✗ Overlook predictions at the individual charging station level.

optimization utilizing an SoC forecasted through the XGBoost performance, e.g., for modeling and prediction tasks. This is algorithm. The data utilized in this study were acquired from critical for IoEV-related applications that involve complex (e.g., the Shanghai New Energy Electric Vehicle Monitoring Center high-dimensional and nonlinear) patterns and dependencies. [206], pertaining to a cohort of 1,000 EVs subjected to testing Two representative techniques are GDM and transformer.

The over the course of one month. The data attributes included former captures inherent distribution characteristics and models speed, acceleration, SoC, temperature, longitudes, and latitudes. complex relationships to understand system dynamics. The The data points were sampled every 10 seconds.

The output latter extracts multi-scale features and exploits long-term variables were the 24-hour SoC predictions for individual dependencies well with its self-attention mechanism. vehicles and the forecast of regional spatial–temporal charging With these strengths, GenAI becomes a versatile and demand. Four categories of vehicle’s SoC, encompassing competitive solution, and we foresee GenAI advancements logistics vehicles, taxis, buses, and private cars, underwent and its increased usage in IoEV applications in the future. predictive analysis.

The assessment criteria exhibited a vari- 2) Selection of GenAI Algorithms: GenAI has been used for ability spanning approximately 1.66% to 3.15% for MAE and various IoEV applications and the optimal GenAI performance 2.32% to 4.44% for RMSE [40]. Future research will refine in part depends on the choice of GenAI algorithms. Each GenAI the findings of this study.

Considering additional factors such algorithm owns unique characteristics that can influence the as road conditions and user demographics will enhance the algorithm’s performance in solving different problems. For accuracy of EV SoC predictions. example, GAN has been shown to be a popular technology for Table VI summarizes GenAI implementations for IoEV data augmentation by generating high-fidelity synthetic data applications across multiple layers.

We discovered that a few (e.g., charging behavior). When robustness is a key concern, papers considered both layer 1 and layer 3 applications. For GAN becomes less competitive due to its ineffectiveness in example, DiffCharge [61] is able to generate EV charging capturing data diversity well and mode collapse issue.

VAE is scenarios for both battery level and EV charging station level. capable of probabilistic data generation and particularly useful In contrast, in [40], SoC prediction and regional charging load in generating stochastic scenarios. However, its reliance on forecasting were completed by a framework with GAIL. the Gaussian latent space sacrifices data details sometimes. The diffusion model outperforms GAN for producing high- F.

Discussion resolution outputs by iteratively refining noisy data, but its demand for computing resources is significant. Overall, we We have presented technical details of GenAI’s usage in urge a comprehensive evaluation of different GenAI algorithms IoEV in different layers in various aspects and we present our and the selection of suitable algorithms to meet specific discussions about GenAI’s performance below. requirements and constraints in different tasks. 1) Key Features for GenAI’s Competitiveness: We may notice that GenAI is not the first ML technology but manages to outperform traditional ML algorithms for many discussed IV.

T ECHNICAL R EVIEWS : DATASET applications and tasks. We summarize and highlight the key Data is highly important for GenAI for model training, features that drive GenAI’s competitiveness in IoEV. system customization, performance improvement, and so on. a) Data Generation and Robustness: Data scarcity has In this section, we provide a summary of the available been among the toughest challenges for traditional ML, e.g., public dataset in the domain of GenAI-based electric mobility in layers 2 and 3 for predicting supply and demand.

GenAI applications. We summarize the datasets in Table VII and algorithms such as GAN and VAE help generate realistic describe them in detail as below. synthetic data by learning the underlying distribution of the Layer 1: In the battery layer, the dataset from NSMC-EV limited raw data. Such GenAI algorithms demonstrate a high [74] and the unlabeled dataset on multiple faults and failure level of noise control during data generation and this enhances scenarios [76] are used by GRU-VAE [51] and BERTtery [25], the stability of anomaly detection, predictive modeling, etc. respectively, for the anomaly detection.

NSMC-EV serves as b) Advanced Pattern Recognition: Compared to tra- China’s national big data platform for EVs, offering extensive ditional ML, GenAI exhibits improved pattern recognition real-time online data on EVs utilized in public transportation.

TABLE VII

S UMMARY OF DATASET USED IN G ENAI FOR I O EV AT DIFFERENT LAYERS .

Applications Dataset Properties Dataset Size Algorithms

Layer 1 NSMC-EV 13-dimensional time series, e.g., vehicle speed, NSMC-EV platform for over GRU-VAE [51] Anomaly [74] charging state, insulation resistance, and SoC. three million EVs Detection Faults and Multiple scenarios, e.g., short circuit and thermal 316 NCM battery cells [25] BERTtery [25] Failure [76] runaway. Time series of voltage, current, etc. Time series, e.g., vehicle speed, voltage, cell Driving data from five identical EV [88] TS-WGAN [65] SoC temperature, motor controller voltage, and SoC. vehicles over a year [88] Estimation Li-ion Battery Time series of recorded variables, e.g., cell voltage, Various drive cycles, including TS-WGAN [65] [89] current, battery temperature, and ampere-hours.

US06, UDDS, and LA92. SoH LFP Battery Rated capacity, number of cells, charging current, 4 cells; cycles: 1062, 1266, DDPM [60] Estimation [105] discharging current, cut-off voltage, etc. 1114, and 1047. [60] Layer 2 EV Charging EA Technology Residential charging events, e.g., date time, arrival Charging behaviors of over 200 GAN [17] Behaviors [115] hour, plug-out hour, and required energy participants in 2014 and 2015. iHomeLab Energy consumption of households and specific Five houses in the Lucerne VAE-GAN [108], Smart Home PART [123] appliances as well as PV generation. region, Switzerland.

GANs [191] CLNR TC1a UK electricity customers’ electricity consumption Up to 8,000 customers for the GAN [119] Residential EV [120] measured by British Gas’s smart meter; year 2011. Load CLNR TC5 TC5 [121]: including customers’ energy use and Part of the project involving GAN [119] [121] solar PV performance. over 12,000 consumers. Layer 3 EV Load City of Charging load records, e.g., address, arrival and 4 years; 25 public charging Transformer [15] Forecasting Boulder [207] departure time, type of the plug, data, and energy. stations in Boulder.

ACN Data Time of EV connection, charging time, amount of Over 30,000 charging sessions; CopulaGAN [137]; Scenarios [153] energy received by EV, and time of leaving. growing daily [153]. DiffCharge [61] Generation Belgian Elia 2 months of 15-minute interval PV and load power in the microgrid. GAN [38] Group [208] data.

Layer 4 Adversarial California Electricity prices and the grid load information Continuously updating [185] RL-AdvGAN [39] Attacks OASIS [185] from OASIS site. ERCOT’s Data FDIAs Load profiles from the grid. Continuously updating [194].

GAE [45] [194] OTIDS dataset States of DoS attacks, fuzzy attacks, impersonation Attacks: 657K DoS, 592K fuzzy, RX-ADS [47] DoS and Fuzzy [195] attacks, and attack-free conditions. and 995K impersonation. Attacks Car-Hacking Attack types, e.g., DoS, fuzzy, drive gear spoofing, Attacks: 3.6M DoS, 3.8M fuzzy, RX-ADS [47] [196] and revolutions per minute (RPM) gauge spoofing. 4.4M spoofing drive gear, etc. Multiple Layers SoC and Load Shanghai New Data sampled for speed, acceleration, SoC, GAIL, PPO, Continuously updating [206] Forecasting Energy [206] temperature, longitudes, and latitudes.

XGBoost [40]

The dataset used for EV’s SoC estimation can be found in [88], from February 2014 to November 2015. EA Technology offers [89], which provides the battery’s voltage, current, temperature, specialized asset management solutions for electrical asset SoC, etc. EV dataset from [88] consists of driving data of owners and operators worldwide [115].

This data is used by [17] 5 identical vehicles over a year where each vehicle’s data for the study of EV charging behaviors. The iHomeLab PART for charging and discharging events is sampled at 10-second dataset [123] includes the energy consumption of households intervals. An LPF battery dataset from [105] is used for EV’s and specific appliances as well as PV generation.

The data are SoH estimation. It consists of 4 cells with a number of charge- collected from 5 houses in the Lucerne region, Switzerland, discharge cycles over a thousand times for each cell. and recorded over 1.5 to 3.5 years. It is used by [108] for data augmentation in the smart home applications, and by [191] for Layer 2: For the EV layer in Table VII, data from EA the study of adversarial attacks.

CLNR’s dataset TC1a [120] Technology [115] consists of residential EV charging events and TC5 [121] consist of UK electricity customers’ overall with charging behaviors of over 200 participants observed

electricity consumption and customers’ energy use and solar 2) Hallucination: Another challenge could be solving the PV performance, respectively. These datasets are used in GAN hallucination issues in GenAI models. The hallucination refers [119] for generating residential EV load.

CLNR is a project to ML generating outputs that are plausible but incorrect. For supported by the Ofgem’s Low Carbon Network Fund, which future work, developing methods that can detect and mitigate aimed to facilitate UK’s low carbon energy sector [209]. hallucinations in GenAI is very important, especially for high- Layer 3: This layer is about the interaction between EVs stakes applications such as EV routing and EV’s battery and power grid.

The City of Boulder Open Data Hub [207] is management system. The hybrid systems combining GenAI used by a Transformer-based model for EV load forecasting. with the traditional rule-based approach or the new architectures It provides EV charging load records spanning about 4 years that can verify the correctness of generated content/data could collected from 25 public charging stations in Boulder, Colorado be developed to avoid hallucination issues. where the charging stations are equipped with 22 kW-rated 3) Transfer Learning: Additionally, models, though are connectors.

Datasets from [208] and [153] are employed for accurate for certain applications, may suffer from performance scenario generations. The Belgian grid dataset from Elia Group drops when the applications are different, even slightly. Transfer [208] includes PV and load in the microgrid with about two learning could be used to apply knowledge gained from one months of data.

ACN data [153] include the time of EV domain (e.g., load forecasting in Europe) to another (e.g., connection, done charging time, amount of energy received load forecasting in Asia), which reduces training time, lowers by EV, and time of EV leaving the parking lot. This dataset computational costs, and improves results with smaller datasets. contains charging sessions at Caltech parking lots in California, 4) Integration of EVs as Distributed Energy Resources: and is continuously growing every day [153].

Furthermore, combining traditional grid-to-vehicle (G2V) and Layer 4: For the security layer, California ISO OASIS site advanced V2G techniques allows EVs act as mobile energy [185] provides the continuously updating dataset for electricity storage units and enables bidirectional energy flow between prices and the grid load information where OASIS offers real- the vehicle and the grid. The G2V techniques view EVs as the time data related to the ISO transmission system and its market. energy consumers, while V2G techniques provide EVs with the Data from [185] is used by RL-AdvGAN [39] for the study opportunity to deliver power back to the grid.

This bidirectional of adversarial attacks. The load profiles from the grid can be energy flow capability transforms EVs into essential elements found in ERCOT [194]. However, it requires an IP address of distributed energy storage systems, provides the benefits from the U.S. to access data from ERCOT.

This dataset is used to the grid such as helping for load balance, mitigating peak by GAE [45] for the investigation of FDIAs. OTIDS dataset demand, enhancing grid resilience and stability. For future [195] and Car-Hacking dataset [196] are employed in [47] for work, developing GenAI-based optimization algorithms which the detection of DoS, fuzzy, and impersonation attacks. can dynamically manage energy flow between EVs and the grid, Multiple Layers: Moreover, datasets can be used in multiple as well as energy transactions between EVs, in coordination layers.

For example, the Shanghai New Energy EV Monitoring with renewable energy sources, is crucial for future of smart Center [206] is a data sharing and cooperation platform with grids and sustainable energy solutions. extensive operational data on new energy vehicles in Shanghai, B. New Methodologies and provides the EV’s data such as speed, acceleration, battery status, and location. The data is continuously updated and used Besides enhancing existing models, new technologies can in [40] for the prediction of EV’s SoC and load. also be explored. 1) LLMs: LLMs are often discussed together with GenAI, both representing the latest ML technology advancements.

V. F UTURE D IRECTIONS LLMs were originally developed for language-centric appli- cations and language has not been the focus of EV related GenAI has demonstrated its potential in the IoEV, and there applications. With the ever-increasing interactions between EV are still areas to be improved.

This section outlines several systems and users, there is a demand to incorporate NLP into directions to unlock new opportunities in the IoEV ecosystem. the technology stack to facilitate the interaction and contribute to the enhancement of IoEV. For example, LLMs such as GPT-4 and LLaMA can offer enhanced interactions between A. Improving Existing Solutions EV users and charging infrastructure.

The EV users can specify their charging preferences in natural language, such as “I need While GenAI models have been developed for electric to charge my car fully by 7 AM, tomorrow morning”. Then, mobility, they are not perfect. the LLMs-based system can translate this into an optimized 1) Up-to-date Models: One challenge is to make sure charging schedule. Such interaction simplifies the process and models are updated and data plays an important role.

Data offers a more accessible and personalized experience for the patterns may evolve over time due to the dynamic nature EV users, not necessarily from a technical background of of IoEV environments. This requires the models to have IoEV. LLMs can also be used to interpret users’ historical continuous learning capability where new information should queries, largely language-based, to predict user demand and be adapted without forgetting previous knowledge.

Integrating optimize real-time responses. Overall, the integration of LLMs continual learning into the GenAI system can keep it remain in IoEV charging systems shall be investigated to improve user effective over the lifespan of an EV. experience and facilitate the adoption of electric mobility.

2) Federated Learning: Privacy concerns, systems’ robust- R EFERENCES ness, and decentralized systems are critical in IoEV applications. [1] S. Becker, “Are US EV sales a disaster or a Federated learning can address security issues by enabling EVs booming segment? the answer may be both,” BBC: to learn collaboratively without sharing sensitive data. Future https://www.bbc.com/worklife/article/20240228-are-us-ev-sales-a- work could study how federated learning can be effectively disaster-or-a-booming-segment-the-answer-may-be-both, 29 Feb. 2024, accessed: 20 Mar 2024. implemented in large-scale IoEV to promote the system’s [2] Editor OilPrice.com, “China’s ev growth set to explode in 2024,” scalability and security.

Yahoo!finance: https://finance.yahoo.com/news/china-ev-growth-set- 3) Hybrid Models: When one model is insufficient to explode-150000597.html, 21 Feb. 2024, accessed: 03 Apr 2024. perform well, the hybrid models can be considered and the [3] VANTAGE Market Research, “Electric vehicle market - global models are not limited to GenAI. For example, combining industry assessment & forecast,” VANTAGE Market Research’s GenAI with DRL may improve the models’ performance in report: https://www.vantagemarketresearch.com/industry-report/ complex IoEV applications.

Several existing research efforts electric-vehicle-ev-market-1853, Feb. 2024, accessed: 03 Apr 2024. [4] S. Zhang, S. Zhang, L.

K. Yeung et al., “Urban internet of electric ve- are GRU and VAE [51] for battery anomaly detection, LSTM hicle parking system for vehicle-to-grid scheduling: Formulation and and GAN for SoC estimation [66], DNN and GAN for EV distributed algorithm,” IEEE Transactions on Vehicular Technology, charging behaviors prediction [17], etc. Future research may 2023. explore cross-layer optimizations with hybrid models, e.g., [5] O.

Gul M. Khan, F. Elghitani, A.

Youssef et al., “Real-time advanced BMS informs routing decisions and grid interactions. congestion-aware charging station assignment model for evs,” IEEE This helps create an integrated and efficient system. Internet of Things Journal, vol. 11, no. 7, pp. 11 723–11 736, 2024. 4) Embodied AI and AGI: Looking toward future ad- [6] H. Wang, S.

Chen, M. Li et al., “Demand-driven charging strategy based distributed routing optimization under traffic restrictions in vancements, the embodied AI and AGI are expected to play internet of electric vehicles,” IEEE Internet of Things Journal, pp. transformative roles in IoEV. The former combines cognitive 1–1, 2024. processing with physical interaction to enhance adaptability [7] S.

Lee and D.-H. Choi, “Three-stage deep reinforcement learning and real-world decision-making, e.g., navigating complex for privacy-and safety-aware smart electric vehicle charging station environments and responding to traffic conditions. The latter scheduling and volt/var control,” IEEE Internet of Things Journal, vol. 11, no. 5, pp. 8578–8589, 2024. aims at achieving broad cognitive abilities similar to human [8] S.

Liu, C. Wu, J. Huang et al., “Blockchain-based interpretable intelligence.

It can potentially revolutionize IoEV, e.g., make electric vehicle battery life prediction in iov,” IEEE Internet of decisions in unpredictable environments without relying on Things Journal, vol. 11, no. 4, pp. 7214–7227, 2024. pre-programmed instructions. [9] M. Naik, A. P.

Singh, N. R. Pradhan et al., “Tokengreen: A versatile Together, new methodologies and technologies could drive nft framework for peer-to-peer energy trading and asset ownership further innovation in IoEV in various aspects such as sustain- of electric vehicles,” IEEE Internet of Things Journal, vol. 11, no. 8, pp. 13 636–13 646, 2024. ability and safety. [10] A.

Mohammad, R. Zamora, and T. T.

Lie, “Transactive energy management of pv-based ev integrated parking lots,” IEEE Systems VI. C ONCLUSIONS Journal, vol. 15, no. 4, pp. 5674–5682, 2021. In this survey, we explored the applications of GenAI [11] J.

Liu, G. Lin, S. Huang et al., “Optimal ev charging scheduling by considering the limited number of chargers,” IEEE Transactions on in the IoEV from various perspectives.

We grouped relevant Transportation Electrification, vol. 7, no. 3, pp. 1112–1122, 2021. applications across four different layers: EV’s battery layer, [12] Y. Li, S. He, Y.

Li et al., “Probabilistic charging power forecast individual EV layer, the grid layer, and security layer. We of evcs: Reinforcement learning assisted deep learning approach,” concluded that these applications primarily utilize GenAI’s IEEE Transactions on Intelligent Vehicles, vol. 8, no. 1, pp. 344–357, capabilities in data feature extraction, enhancement, and 2023. generation. The characteristics of GenAI make it ideal for [13] X.

Diao, L. Jiang, T. Gao et al., “Research on electric vehicle charging safety warning based on a-lstm algorithm,” IEEE Access, data augmentation.

Hence, research works in the first three vol. 11, pp. 55 081–55 093, 2023. layers all used GenAI for this purpose. At layer 1, GenAIs [14] Z. Pan, J.

Wang, W. Liao et al., “Data-driven ev load profiles including GAN, GDM, VAE, and Transformer were introduced generation using a variational auto-encoder,” Energies, vol. 12, no. 5, for anomaly detection, as well as SoC and SoH estimations 2019. of the EV’s battery. At layer 2, GAN and Transformer were [15] S.

Koohfar, W. Woldemariam, and A. Kumar, “Prediction of electric discussed for EV charging behaviors and loads as well as vehicles charging demand: A transformer-based deep learning the optimal EV routing problem.

At layer 3, GAN, VAE, and approach,” Sustainability, vol. 15, no. 3, 2023. [16] X. Shen, H. Zhao, Y.

Xiang et al., “Short-term electric vehicles Transformer are the main techniques currently employed for EV charging load forecasting based on deep learning in low-quality charging load forecasting/charging scenarios generation, and data environments,” Electric Power Systems Research, vol. 212, p. LLMs for the analysis of EV charging reviews. At layer 4, GAN 108247, 2022. and AE were often applied for the detection and generation [17] A.

Forootani, M. Rastegar, and H. Zareipour, “Transfer learning- of attacks that may threaten the systems across layers 1 to 3. based framework enhanced by deep generative model for cold-start Subsequently, we summarized the publicly available dataset and forecasting of residential ev charging behavior,” IEEE Transactions on Intelligent Vehicles, pp. 1–9, 2023. the possible further research directions for GenAI’s applications [18] BioLogic, “Battery states: State of charge (SoC), state of health in IoEV.

In conclusion, this survey highlights the essential role (SoH). electrochemistry basics series.” https://www.biologic.net/ of GenAI in IoEV and underscores the urgent need for further topics/battery-states-state-of-charge-soc-state-of-health-soh/, 06 Oct exploration of its applications. 2023, accessed: 05 May 2024.

[19] A. Vaswani, N. Shazeer, N.

Parmar et al., “Attention is all you need,” [41] J. Yoon, J. Jordon, and M. van der Schaar, “GAIN: Missing data Advances in neural information processing systems, vol. 30, 2017. imputation using generative adversarial nets,” in Proceedings of the [20] F.

Liu, B. Ding, C. Guan et al., “Demystify adult learning: A social 35th International Conference on Machine Learning, vol. 80, Jul network and large language model assisted approach,” in 2024 IEEE 2018, pp. 5689–5698.

Annual Congress on Artificial Intelligence of Things (AIoT), 2024, [42] D. Pathak, P. Krahenbuhl, J.

Donahue et al., “Context encoders: Fea- pp. 207–212. ture learning by inpainting,” in Proceedings of the IEEE conference [21] Z. Kwan, W. Zhang, Z.

Wang et al., “Nutrition estimation for on computer vision and pattern recognition, 2016, pp. 2536–2544. dietary management: A transformer approach with depth sensing,” [43] Z. Liu, R. A.

Yeh, X. Tang et al., “Video frame synthesis using deep arXiv:2406.01938, 2024. voxel flow,” in Proceedings of the IEEE international conference on [22] A. Radford, J.

W. Kim, T. Xu et al., “Robust speech recognition via computer vision, 2017, pp. 4463–4471. large-scale weak supervision,” arXiv:2212.04356, 2022. [44] T.

Chen, S. Kornblith, K. Swersky et al., “Big self-supervised models [23] W.

Kim, B. Son, and I. Kim, “ViLT: Vision-and-language transformer are strong semi-supervised learners,” Advances in neural information without convolution or region supervision,” arXiv:2102.03334, 2021. processing systems, vol. 33, pp. 22 243–22 255, 2020. [24] S.

Ha, D. J. Marchetto, S.

Dharur et al., “Topic classification of [45] S. R. Fahim, R.

Atat, C. Kececi et al., “Graph autoencoder-based electric vehicle consumer experiences with transformer-based deep power attacks detection for resilient electrified transportation systems,” learning,” Patterns, vol. 2, no. 2, p. 100195, 2021. IEEE Transactions on Transportation Electrification, pp. 1–1, 2024. [25] J.

Zhao, X. Feng, J. Wang et al., “Battery fault diagnosis and failure [46] H.

S. Mavikumbure, V. Cobilean, C.

S. Wickramasinghe et al., “Cy- prognosis for electric vehicles using spatio-temporal transformer phy ads: Cyber physical anomaly detection framework for ev charging networks,” Applied Energy, vol. 352, p. 121949, 2023. systems,” IEEE Transactions on Transportation Electrification, pp. [26] M. Tang, W.

Zhuang, B. Li et al., “Energy-optimal routing for 1–1, 2024. electric vehicles using deep reinforcement learning with transformer,” [47] C. S.

Wickramasinghe, D. L. Marino, H.

S. Mavikumbure et al., “Rx- Applied Energy, vol. 350, p. 121711, 2023. ads: Interpretable anomaly detection using adversarial ml for electric [27] X. Huang, D.

Wu, and B. Boulet, “Metaprobformer for charging vehicle can data,” IEEE Transactions on Intelligent Transportation load probabilistic forecasting of electric vehicle charging stations,” Systems, vol. 24, no. 12, pp. 14 051–14 063, 2023. IEEE Transactions on Intelligent Transportation Systems, vol. 24, [48] R.

Yeh, Z. Liu, D. B.

Goldman et al., “Semantic facial expression no. 10, pp. 10 445–10 455, 2023. editing using autoencoded flow,” arXiv:1611.09961, 2016. [28] I. Goodfellow, J. Pouget-Abadie, M.

Mirza et al., “Generative [49] J. Walker, C. Doersch, A.

Gupta et al., “An uncertain future: adversarial nets,” Advances in neural information processing systems, Forecasting from static images using variational autoencoders,” in vol. 27, 2014. Computer Vision–ECCV 2016: 14th European Conference, Amster- [29] A. Radford, L.

Metz, and S. Chintala, “Unsupervised representation dam, The Netherlands, October 11–14, 2016, Proceedings, Part VII learning with deep convolutional generative adversarial networks,” 14. Springer, 2016, pp. 835–851. arXiv:1511.06434, 2015. [50] L.

Pan, X. Chen, Z. Cai et al., “Variational relational point completion [30] T.

Karras, T. Aila, S. Laine et al., “Progressive growing of gans for network,” in Proceedings of the IEEE/CVF conference on computer improved quality, stability, and variation,” arXiv:1710.10196, 2017. vision and pattern recognition, 2021, pp. 8524–8533. [31] M.

Arjovsky, S. Chintala, and L. Bottou, “Wasserstein generative [51] C.

Sun, Z. He, H. Lin et al., “Anomaly detection of power battery adversarial networks,” in International conference on machine pack using gated recurrent units based variational autoencoder,” learning.

PMLR, 2017, pp. 214–223. Applied Soft Computing, vol. 132, p. 109903, 2023. [32] T. Karras, S.

Laine, and T. Aila, “A style-based generator architecture [52] J. Ho, A.

Jain, and P. Abbeel, “Denoising diffusion probabilistic for generative adversarial networks,” in Proceedings of the IEEE/CVF models,” Advances in neural information processing systems, vol. 33, conference on computer vision and pattern recognition, 2019, pp. pp. 6840–6851, 2020. 4401–4410. [33] S. Tang, Q.

Yang, D. Gündüz et al., “Evolving semantic communi- [53] J. Song, C.

Meng, and S. Ermon, “Denoising diffusion implicit cation with generative model,” arXiv:2403.20237, 2024. models,” arXiv:2010.02502, 2020. [34] T. Han, J.

Tang, Q. Yang et al., “Generative model based highly [54] R. Rombach, A.

Blattmann, D. Lorenz et al., “High-resolution efficient semantic communication approach for image transmission,” image synthesis with latent diffusion models,” in Proceedings of the in ICASSP 2023-2023 IEEE International Conference on Acoustics, IEEE/CVF conference on computer vision and pattern recognition, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5. 2022, pp. 10 684–10 695. [35] C.-H.

Lee, Z. Liu, L. Wu et al., “Maskgan: Towards diverse [55] L.

Zhang, A. Rao, and M. Agrawala, “Adding conditional control to and interactive facial image manipulation,” in Proceedings of the text-to-image diffusion models,” in Proceedings of the IEEE/CVF IEEE/CVF conference on computer vision and pattern recognition, International Conference on Computer Vision, 2023, pp. 3836–3847. 2020, pp. 5549–5558. [56] H.

Du, R. Zhang, Y. Liu et al., “Enhancing deep reinforcement [36] C.

Hu, F. Cheng, Y. Zhao et al., “State of charge estimation for learning: A tutorial on generative diffusion models in network lithium-ion batteries based on data augmentation with generative optimization,” IEEE Communications Surveys & Tutorials, pp. 1–1, adversarial network,” Journal of Energy Storage, vol. 80, p. 110004, 2024. 2024. [57] J.

Wang, H. Du, D. Niyato et al., “A unified framework for guiding [37] W.

Gang, L. Wu, and G. Xuan, “A load forecasting method of generative ai with wireless perception in resource constrained mobile electric vehicles charging station group based on GAN-RF model,” edge networks,” IEEE Transactions on Mobile Computing, pp. 1–17, in 2021 IEEE 5th Conference on Energy Internet and Energy System 2024.

Integration (EI2), 2021, pp. 3076–3079. [58] ——, “Generative artificial intelligence assisted wireless sensing: [38] W. Shi, D. Zhang, X.

Han et al., “Coordinated operation of active Human flow detection in practical communication environments,” distribution network, networked microgrids, and electric vehicle: A IEEE Journal on Selected Areas in Communications, pp. 1–1, 2024. multi-agent ppo optimization method,” CSEE Journal of Power and [59] ——, “Generative ai for integrated sensing and communication: Energy Systems, pp. 1–12, 2023. Insights from the physical layer perspective,” IEEE Wireless Com- [39] Z.

Peng, Q. Yang, D. Li et al., “Adversarial attacks on deep munications, pp. 1–10, 2024. reinforcement learning applications in electric vehicle charging [60] C.

Luo, Z. Zhang, S. Zhu et al., “State-of-health prediction of scheduling: A dual-stage attack framework,” SSRN, 2023. lithium-ion batteries based on diffusion model with transfer learning,” [40] H.

Li, D. Liang, Y. Zhou et al., “Spatiotemporal charging demand Energies, vol. 16, no. 9, 2023. models for electric vehicles considering user strategies,” Frontiers in Energy Research, vol. 10, 2023.

[61] S. Li, H. Xiong, and Y.

Chen, “Diffcharge: Generating ev charging [80] Z. Zhao, Y. Zhang, Y.

Yang et al., “Load forecasting via grey model- scenarios via a denoising diffusion model,” IEEE Transactions on least squares support vector machine model and spatial-temporal Smart Grid, pp. 1–1, 2024. distribution of electric consumption intensity,” Energy, vol. 255, p. [62] Z. W. Boyang Chen and R.

Zhao, “From fiction to fact: the growing 124468, 2022. role of generative ai in business and finance,” Journal of Chinese [81] X. Fan, W. Zhang, C.

Zhang et al., “Soc estimation of li-ion battery Economic and Business Studies, vol. 21, no. 4, pp. 471–496, 2023. using convolutional neural network with u-net architecture,” Energy, [63] K. Moulaei, A. Yadegari, M.

Baharestani et al., “Generative artificial vol. 256, p. 124612, 2022. intelligence in healthcare: A scoping review on benefits, challenges [82] F. Yang, X. Song, F.

Xu et al., “State-of-charge estimation of lithium- and applications,” International Journal of Medical Informatics, vol. ion batteries via long short-term memory network,” IEEE Access, 188, p. 105474, 2024. vol. 7, pp. 53 792–53 799, 2019. [64] O. Noroozi, S. Soleimani, M.

Farrokhnia et al., “Generative ai in [83] X. Zhang, Y. Qin, C.

Yuen et al., “Time-series regeneration with education: Pedagogical, theoretical, and methodological perspectives.” convolutional recurrent generative adversarial network for remaining International Journal of Technology in Education, vol. 7, no. 3, pp. useful life estimation,” IEEE Transactions on Industrial Informatics, 373–385, 2024. vol. 17, no. 10, pp. 6820–6831, 2021. [65] X. Gu, K. See, Y.

Liu et al., “A time-series wasserstein gan method [84] H. Ni, L. Szpruch, M.

Sabate-Vidales et al., “Sig-wasserstein gans for state-of-charge estimation of lithium-ion batteries,” Journal of for time series generation,” in Proceedings of the Second ACM Power Sources, vol. 581, p. 233472, 2023. International Conference on AI in Finance, ser. ICAIF ’21.

New [66] X. Qiu, S. Wang, and K.

Chen, “A conditional generative adversarial York, NY, USA: Association for Computing Machinery, 2022. network-based synthetic data augmentation technique for battery [85] F. Naaz, J. Channegowda, M.

Lakshminarayanan et al., “Solving state-of-charge estimation,” Applied Soft Computing, vol. 142, p. limited data challenges in battery parameter estimators by using 110281, 2023. generative adversarial networks,” in 2021 IEEE PES/IAS PowerAfrica, [67] MoviTHERM, “Mitigating EV battery fires with infrared technol- 2021, pp. 1–3. ogy,” https://movitherm.com/2024/03/21/blog/mitigating-ev-battery- [86] Y. Tian, Q. Dong, J.

Tian et al., “Capacity estimation of lithium-ion fires-with-infrared-technology/, 21 Mar 2024, accessed: 28 Apr 2024. batteries based on optimized charging voltage section and virtual [68] A. Naha, A. Khandelwal, S.

Agarwal et al., “Internal short circuit sample generation,” Applied Energy, vol. 332, p. 120516, 2023. detection in li-ion batteries using supervised machine learning,” [87] Q. Wang, X. Zhou, C.

Wang et al., “Wgan-based synthetic minority Scientific reports, vol. 10, no. 1, p. 1301, 2020. over-sampling technique: Improving semantic fine-grained classi- [69] J. Xie, L. Zhang, T.

Yao et al., “Quantitative diagnosis of internal fication for lung nodules in ct images,” IEEE Access, vol. 7, pp. short circuit for cylindrical li-ion batteries based on multiclass 18 450–18 463, 2019. relevance vector machine,” Journal of Energy Storage, vol. 32, p. [88] X. Gu, K. See, Y.

Wang et al., “The sliding window and shap 101957, 2020. theory—an improved system with a long short-term memory network [70] Y. Jia, J. Li, C.

Yuan et al., “Data-driven safety risk prediction of model for state of charge prediction in electric vehicle application,” lithium-ion battery,” Advanced Energy Materials, vol. 11, no. 18, p. Energies, vol. 14, no. 12, 2021. 2003868, 2021. [89] C. Vidal, P.

Kollmeyer, E. Chemali et al., “Li-ion battery state of [71] J. Hong, Z.

Wang, and Y. Yao, “Fault prognosis of battery system charge estimation using long short-term memory recurrent neural based on accurate voltage abnormity prognosis using long short-term network with transfer learning,” in 2019 IEEE Transportation memory neural networks,” Applied Energy, vol. 251, p. 113381, Electrification Conference and Expo (ITEC), 2019, pp. 1–6. 2019. [90] TOMO Battery, “Lg inr18650hg2,” Quick Overview: https:// [72] J.

Hong, Z. Wang, W. Chen et al., “Synchronous multi-parameter www.tomobattery.com/product/lg-inr18650hg2/, Apr 2024, accessed: prediction of battery systems on electric vehicles using long short- 27 Apr 2024. term memory networks,” Applied Energy, vol. 254, p. 113648, 2019. [91] Monolithic Power Systems, Inc. , “The importance [73] H.

Zhao, Y. Wang, J. Duan et al., “Multivariate time-series anomaly of state of charge (SOC) and state of health (SOH) detection via graph attention network,” in 2020 IEEE International in battery management systems,” MPS Article: https: Conference on Data Mining (ICDM), 2020, pp. 841–850. //www.monolithicpower.com/en/the-importance-of-state-of-charge- [74] D.

Li, Z. Zhang, P. Liu et al., “Battery fault diagnosis for electric and-state-of-health-in-battery-management-systems, accessed: 28 vehicles based on voltage abnormality by combining the long short- Apr 2024. term memory neural network and the equivalent circuit model,” IEEE [92] X.

Lai, C. Deng, J. Li et al., “Rapid sorting and regrouping of Transactions on Power Electronics, vol. 36, no. 2, pp. 1303–1315, retired lithium-ion battery modules for echelon utilization based on 2021. partial charging curves,” IEEE Transactions on Vehicular Technology, [75] N.

Smol, “The risk of lithium-ion batteries – thermal runaway in vol. 70, no. 2, pp. 1246–1254, 2021. EV’s,” Fire Isolator News: https://fireisolator.com/the-risk-of-lithium- [93] W. Li, D.

Cao, D. Jöst et al., “Parameter sensitivity analysis ion-batteries-thermal-runaway-in-evs/, 28 Jun 2023, accessed: 29 of electrochemical model-based battery management systems for Apr 2024. lithium-ion batteries,” Applied Energy, vol. 269, p. 115104, 2020. [76] J. Zhao, H.

Ling, J. Wang et al., “Data-driven prediction of battery [94] S. Amir, M.

Gulzar, M. O. Tarar et al., “Dynamic equivalent circuit failure for electric vehicles,” Iscience, vol. 25, no. 4, 2022. model to estimate state-of-health of lithium-ion batteries,” IEEE [77] BMS Control Systems Limited, “Exploring BMS state of Access, vol. 10, pp. 18 279–18 288, 2022. charge (SOC): Monitoring battery health,” BMS Controls Arti- [95] H.-T.

Lin, T.-J. Liang, and S.-M. Chen, “Estimation of battery state cles: https://bmscontrols.co.uk/blog/exploring-bms-state-of-charge- of health using probabilistic neural network,” IEEE Transactions on soc-monitoring-battery-health/, accessed: 28 Apr 2024.

Industrial Informatics, vol. 9, no. 2, pp. 679–685, 2013. [78] K. S. Mawonou, A.

Eddahech, D. Dumur et al., “State-of-health [96] Z. Deng, L.

Yang, Y. Cai et al., “Online available capacity prediction estimators coupled to a random forest approach for lithium-ion and state of charge estimation based on advanced data-driven battery aging factor ranking,” Journal of Power Sources, vol. 484, p. algorithms for lithium iron phosphate battery,” Energy, vol. 112, 229154, 2021. pp. 469–480, 2016. [79] J. Wang, R.

Zhao, Q.-A. Huang et al., “High-efficient prediction of [97] C. Luo, Z.

Zhang, D. Qiao et al., “Life prediction under charging state of health for lithium-ion battery based on ac impedance feature process of lithium-ion batteries based on automl,” Energies, vol. 15, tuned with gaussian process regression,” Journal of Power Sources, no. 13, 2022. vol. 561, p. 232737, 2023. [98] X. Li, Z.

Wang, and J. Yan, “Prognostic health condition for lithium battery using the partial incremental capacity and gaussian process regression,” Journal of Power Sources, vol. 421, pp. 56–67, 2019.

[99] S. Sohn, H.-E. Byun, and J.

H. Lee, “Cnn-based online diagnosis of [119] F. Charbonnier, T.

Morstyn, and M. McCulloch, “Home electricity knee-point in li-ion battery capacity fade curve,” IFAC-PapersOnLine, data generator (hedge): An open-access tool for the generation of vol. 55, no. 7, pp. 181–185, 2022, 13th IFAC Symposium on electric vehicle, residential demand, and pv generation profiles,” Dynamics and Control of Process Systems, including Biosystems MethodsX, vol. 12, p. 102618, 2024. DYCOPS 2022. [120] Customer-Led Network Revolution, “Dataset (tc1a): [100] Y.

Harada, L. Lin, and M. Fukui, “An rnn approach for lithium-ion Basic profiling of domestic smart meter customers,” battery internal impedance estimation,” in 2020 IEEE REGION 10 http://www.networkrevolution.co.uk/project-library/dataset-tc1a- CONFERENCE (TENCON).

IEEE, 2020, pp. 1145–1146. basic-profiling-domestic-smart-meter-customers/, 28 Aug 2014, [101] M. Singh, S. Bansal, Vandana et al., “A Genetic Algorithm and RNN- accessed: 16 Jul 2024.

LSTM Model for Remaining Battery Capacity Prediction,” Journal [121] ——, “Dataset (tc5): Enhanced profiling of domestic customers of Computing and Information Science in Engineering, vol. 22, no. 4, with solar photovoltaics (pv),” http://www.networkrevolution.co.uk/ p. 041009, 02 2022. project-library/dataset-tc5-enhanced-profiling-solar-photovoltaic- [102] R. Rouhi Ardeshiri and C. Ma, “Multivariate gated recurrent unit for pv-users/, 29 Aug 2014, accessed: 16 Jul 2024. battery remaining useful life prediction: A deep learning approach,” [122] Department for Transport (2021), “National travel survey, 2002- International Journal of Energy Research, vol. 45, no. 11, pp. 16 633– 2020,” http://doi.org/10.5255/UKDA-SN-5340-11, 25 October 2021, 16 648, 2021. accessed: 17 Jul 2024. [103] Y.

Zhao, Z. Wei, Q. Yan et al., “Practical battery health monitoring [123] P.

Huber, M. Ott, M. Friedli et al., “Residential power traces for five using uncertainty-aware bayesian neural network,” arXiv:2404.14444, houses: the ihomelab rapt dataset,” Data, vol. 5, no. 1, p. 17, 2020. 2024. [124] Y.

Yingfei, Z. Mengze, L. Zeyu et al., “Green logistics performance [104] L.

Yang, Z. Zhang, Y. Song et al., “Diffusion models: A comprehen- and infrastructure on service trade and environment-measuring firm’s sive survey of methods and applications,” ACM Computing Surveys, performance and service quality,” Journal of King Saud University - vol. 56, no. 4, pp. 1–39, 2023.

Science, vol. 34, no. 1, p. 101683, 2022. [105] K. A. Severson, P.

M. Attia, N. Jin et al., “Data-driven prediction of battery cycle life before capacity degradation,” Nature Energy, [125] O.

Vinyals, M. Fortunato, and N. Jaitly, “Pointer networks,” Advances vol. 4, no. 5, pp. 383–391, 2019. in neural information processing systems, vol. 28, 2015. [106] D.

Chen, W. Hong, and X. Zhou, “Transformer network for remaining [126] T.

Barrett, W. Clements, J. Foerster et al., “Exploratory combinatorial useful life prediction of lithium-ion batteries,” IEEE Access, vol. 10, optimization with reinforcement learning,” in Proceedings of the pp. 19 621–19 628, 2022.

AAAI conference on artificial intelligence, vol. 34, 2020, pp. 3243– [107] S. Li, X. Jin, Y.

Xuan et al., “Enhancing the locality and breaking 3250. the memory bottleneck of transformer on time series forecasting,” [127] W. Pan and S. Q.

Liu, “Deep reinforcement learning for the dynamic Advances in neural information processing systems, vol. 32, 2019. and uncertain vehicle routing problem,” Applied Intelligence, vol. 53, [108] M. Razghandi, H. Zhou, M.

Erol-Kantarci et al., “Smart home energy no. 1, pp. 405–422, 2023. management: Vae-gan synthetic dataset generator and q-learning,” [128] W. Kool, H. Van Hoof, and M.

Welling, “Attention, learn to solve IEEE Transactions on Smart Grid, vol. 15, no. 2, pp. 1562–1573, routing problems!” arXiv:1803.08475, 2018. 2024. [129] S. Erdoğan and E. Miller-Hooks, “A green vehicle routing problem,” [109] Z.

Chen, Z. Zhang, J. Zhao et al., “An analysis of the charging Transportation Research Part E: Logistics and Transportation Review, characteristics of electric vehicles based on measured data and its vol. 48, no. 1, pp. 100–114, 2012, select Papers from the 19th application,” IEEE Access, vol. 6, pp. 24 475–24 487, 2018.

International Symposium on Transportation and Traffic Theory. [110] H. Jahangir, H. Tayarani, A.

Ahmadian et al., “Charging demand [130] Gurobi Optimization, LLC, “Computer Software: Gurobi Optimizer of plug-in electric vehicles: Forecasting travel behavior based on a (Version 9.5),” https://www.gurobi.com, 2022. novel rough artificial neural network approach,” Journal of cleaner [131] S. Zhang, Y. Gajpal, S.

Appadoo et al., “Electric vehicle routing production, vol. 229, pp. 1029–1044, 2019. problem with recharging stations for minimizing energy consumption,” [111] X. Zhang, K. W.

Chan, H. Li et al., “Deep-learning-based proba- International journal of production economics, vol. 203, pp. 404–413, bilistic forecasting of electric vehicle charging load with a novel 2018. queuing model,” IEEE transactions on cybernetics, vol. 51, no. 6, [132] D. Goeke and M.

Schneider, “Routing a mixed fleet of electric and pp. 3157–3170, 2020. conventional vehicles,” European Journal of Operational Research, [112] A. J. Jahromi, M.

Mohammadi, S. Afrasiabi et al., “Probability vol. 245, no. 1, pp. 81–99, 2015. density function forecasting of residential electric vehicles charging [133] R. Zhang, K.

Xiong, H. Du et al., “Generative ai-enabled vehicular profile,” Applied Energy, vol. 323, p. 119616, 2022. networks: Fundamentals, framework, and case study,” IEEE Network, [113] A. Forootani, M.

Rastegar, and A. Sami, “Short-term individual vol. 38, no. 4, pp. 259–267, 2024. residential load forecasting using an enhanced machine learning- [134] M. Andreoni, W.

T. Lunardi, G. Lawton et al., “Enhancing au- based approach based on a feature engineering framework: A tonomous system security and resilience with generative ai: A comparative study with deep learning methods,” Electric Power comprehensive survey,” IEEE Access, vol. 12, pp. 109 470–109 493, Systems Research, vol. 210, p. 108119, 2022. 2024. [114] M.

Gilanifar and M. Parvania, “Clustered multi-node learning of electric vehicle charging flexibility,” Applied energy, vol. 282, p. [135] H. Lin, Y.

Liu, S. Li et al., “How generative adversarial networks 116125, 2021. promote the development of intelligent transportation systems: A survey,” IEEE/CAA Journal of Automatica Sinica, vol. 10, no. 9, pp. [115] EA Technology, “My electric avenue data download,” MPS Arti- 1781–1796, 2023. cle: https://eatechnology.com/resources/projects/my-electric-avenue- data-download/, 04 May 2017, accessed: 02 May 2024. [136] N. Huang, Q.

He, J. Qi et al., “Multinodes interval electric vehicle [116] Y.-W. Chung, B.

Khaki, T. Li et al., “Ensemble machine learning- day-ahead charging load forecasting based on joint adversarial based algorithm for electric vehicle user behavior prediction,” Applied generation,” International Journal of Electrical Power & Energy Energy, vol. 254, p. 113732, 2019. Systems, vol. 143, p. 108404, 2022. [117] S.

Shahriar, A.-R. Al-Ali, A. H.

Osman et al., “Prediction of ev [137] M. A. Gharibi, H.

Nafisi, H. Askarian-abyaneh et al., “Deep learning charging behavior using machine learning,” Ieee Access, vol. 9, pp. framework for day-ahead optimal charging scheduling of electric 111 576–111 586, 2021. vehicles in parking lot,” Applied Energy, vol. 349, p. 121614, 2023. [118] S. Zhang, Z.

Ji, J. Zhang et al., “Multi-dimensional data generation method of electric vehicle charging behaviors based on improved generative adversarial network,” in 2022 IEEE / IAS Industrial and Commercial Power System Asia (I & CPS Asia), 2022, pp. 1827– 1832.

[138] Y. Luo, X. Cai, Y.

Zhang et al., “Multivariate time series impu- [158] J. Devlin, M.-W. Chang, K.

Lee et al., “Bert: Pre-training tation with generative adversarial networks,” Advances in neural of deep bidirectional transformers for language understanding,” information processing systems, vol. 31, 2018. arXiv:1810.04805, 2018. [139] M. Gilanifar and M. Parvania, “Clustered multi-node learning of [159] Z.

Yang, Z. Dai, Y. Yang et al., “Xlnet: Generalized autoregres- electric vehicle charging flexibility,” Applied energy, vol. 282, p. sive pretraining for language understanding,” Advances in neural 116125, 2021. information processing systems, vol. 32, 2019. [140] R.

Buechler, E. Balogun, A. Majumdar et al., “Evgen: Adversarial [160] K.

Alvarez, “Evaluating electric vehicle user mobility data using networks for learning electric vehicle charging loads and hidden neural network-based language models,” in Proceedings of the representations,” arXiv:2108.03762, 2021. Transportation Research Board (TRB) annual meeting, 2019. [141] M. Jamshidian and M.

Mata, “2 - advances in analysis of mean [161] O. F. Zaidan, J.

Eisner, and C. Piatko, “Machine learning with and covariance structure when data are incomplete,” in Handbook annotator rationales to reduce annotation cost,” in Proceedings of the of Latent Variable and Related Models, 2007, pp. 21–44. NIPS* 2008 workshop on cost sensitive learning, 2008, pp. 260–267. [142] Tobin A.

Driscoll and Richard J. Braun, “5.2. piecewise linear [162] S. Serrano and N.

A. Smith, “Is attention interpretable?” interpolation — fundamentals of numerical computation,” https: arXiv:1906.03731, 2019. //tobydriscoll.net/fnc-julia/localapprox/pwlin.html, 2022, accessed: [163] A. T.

Nguyen, B. C. Wallace, J.

J. Li et al., “Aggregating and pre- 19 May 2024. dicting sequence labels from crowd annotations,” in Proceedings of [143] S. Zhang, “Nearest neighbor selection for iteratively knn imputation,” the conference.

Association for Computational Linguistics. Meeting, Journal of Systems and Software, vol. 85, no. 11, pp. 2541–2552, vol. 2017. NIH Public Access, 2017, p. 299. 2012. [164] J.

Qian, Y. Jiang, X. Liu et al., “Federated reinforcement learning for [144] Q.

Sun, J. Liu, X. Rong et al., “Charging load forecasting of electric electric vehicles charging control on distribution networks,” IEEE vehicle charging station based on support vector regression,” in 2016 Internet of Things Journal, vol. 11, no. 3, pp. 5511–5525, 2024.

IEEE PES Asia-Pacific Power and Energy Engineering Conference [165] F. Zhang, Q. Yang, and D.

An, “Cddpg: A deep-reinforcement- (APPEEC). IEEE, 2016, pp. 1777–1781. learning-based approach for electric vehicle charging control,” IEEE [145] T. S.

Mahmoud, D. Habibi, M. Y.

Hassan et al., “Modelling self- Internet of Things Journal, vol. 8, no. 5, pp. 3075–3087, 2021. optimised short term load forecasting for medium voltage loads [166] H. Li, Z. Wan, and H.

He, “Constrained EV charging scheduling using tunning fuzzy systems and artificial neural networks,” Energy based on safe deep reinforcement learning,” IEEE Transactions on Conversion and Management, vol. 106, pp. 1396–1408, 2015. Smart Grid, vol. 11, no. 3, pp. 2427–2439, 2020. [146] G. Lai, W.-C.

Chang, Y. Yang et al., “Modeling long-and short- [167] J. Zhang, L.

Che, and M. Shahidehpour, “Distributed training and term temporal patterns with deep neural networks,” in The 41st distributed execution-based Stackelberg multi-agent reinforcement international ACM SIGIR conference on research & development in learning for EV charging scheduling,” IEEE Transactions on Smart information retrieval, 2018, pp. 95–104. Grid, vol. 14, no. 6, pp. 4976–4979, 2023. [147] X.

Huang, D. Wu, and B. Boulet, “Ensemble learning for charging load forecasting of electric vehicle charging stations,” in 2020 IEEE [168] J.

Jin, S. Mao, and Y. Xu, “Optimal priority rule-enhanced deep Electric Power and Energy Conference (EPEC).

IEEE, 2020, pp. reinforcement learning for charging scheduling in an electric vehicle 1–5. battery swapping station,” IEEE Transactions on Smart Grid, vol. 14, no. 6, pp. 4581–4593, 2023. [148] M. Li and M. Lenzen, “How many electric vehicles can the current australian electricity grid support?” International Journal of [169] J.

Wei, L. Yao, and Q. Meng, “Self-adaptive logit balancing for deep Electrical Power & Energy Systems, vol. 117, p. 105586, 2020. neural network robustness: Defence and detection of adversarial attacks,” Neurocomputing, vol. 531, pp. 180–194, 2023. [149] N.

Nguyen and B. Quanz, “Temporal latent auto-encoder: A method for probabilistic multivariate time series forecasting,” Proceedings [170] A. Chakraborty, M.

Alam, V. Dey et al., “A survey on adversarial of the AAAI Conference on Artificial Intelligence, vol. 35, no. 10, attacks and defences,” CAAI Transactions on Intelligence Technology, pp. 9117–9125, May 2021. vol. 6, no. 1, pp. 25–45, 2021. [150] H. Zhou, S.

Zhang, J. Peng et al., “Informer: Beyond efficient [171] S. Huang, N.

Papernot, I. Goodfellow et al., “Adversarial attacks on transformer for long sequence time-series forecasting,” Proceedings neural network policies,” arXiv:1702.02284, 2017. of the AAAI Conference on Artificial Intelligence, vol. 35, no. 12, [172] A. Kurakin, I.

J. Goodfellow, and S. Bengio, “Adversarial examples pp. 11 106–11 115, 2021. in the physical world,” in Artificial intelligence safety and security. [151] I.

Svetunkov, “Forecasting and analytics with the augmented Chapman and Hall/CRC, 2018, pp. 99–112. dynamic adaptive model (adam),” https://openforecast.org/adam/ [173] S.-M. Moosavi-Dezfooli, A. Fawzi, and P.

Frossard, “Deepfool: typesOfForecasts.html, 08 May 2024, accessed: 20 May 2024. a simple and accurate method to fool deep neural networks,” in [152] A. Nichol, “On first-order meta-learning algorithms,” Proceedings of the IEEE conference on computer vision and pattern arXiv:1803.02999, 2018. recognition, 2016, pp. 2574–2582. [153] Z. J.

Lee, T. Li, and S. H.

Low, “Acn-data: Analysis and applications [174] P. Wang, Y. Li, S.

Shekhar et al., “Adversarial attacks on reinforce- of an open ev charging dataset,” in Proceedings of the Tenth ACM ment learning based energy management systems of extended range International Conference on Future Energy Systems, ser. e-Energy electric delivery vehicles,” arXiv:2006.00817, 2020. ’19. New York, NY, USA: Association for Computing Machinery, [175] W. Gan, M.

Shahidehpour, J. Guo et al., “A tri-level planning 2019, p. 139–149. approach to resilient expansion and hardening of coupled power [154] Bundesnetzagentur, “Smard platform,” https://www.smard.de/en, distribution and transportation systems,” IEEE Transactions on Power 2024, accessed: 29 May 2024. Systems, vol. 37, no. 2, pp. 1495–1507, 2021. [155] United States Environmental Protection Agency, “Inventory [176] K.-D.

Lu and Z.-G. Wu, “Multi-objective false data injection attacks of u.s. greenhouse gas emissions and sinks: 1990-2016,” of cyber–physical power systems,” IEEE Transactions on Circuits https://www.epa.gov/ghgemissions/inventory-us-greenhouse- and Systems II: Express Briefs, vol. 69, no. 9, pp. 3924–3928, 2022. gas-emissions-and-sinks-1990-2016, 2018, accessed: 14 Jun 2024. [177] Y. Li and Y.

Wang, “False data injection attacks with incomplete [156] N. R. Council, G.

Affairs, B. on Science et al., Hidden costs of energy: network topology information in smart grid,” IEEE Access, vol. 7, unpriced consequences of energy production and use. National pp. 3656–3664, 2018. Academies Press, 2010. [178] D.

Xue, X. Jing, and H. Liu, “Detection of false data injection [157] O.

I. Asensio, K. Alvarez, A.

Dror et al., “Real-time data from attacks in smart grid utilizing elm-based ocon framework,” IEEE mobile platforms to evaluate sustainable transportation infrastructure,” Access, vol. 7, pp. 31 762–31 773, 2019. Nature Sustainability, vol. 3, no. 6, pp. 463–471, 2020.

[179] O. Boyaci, A. Umunnakwe, A.

Sahu et al., “Graph neural networks [195] Hacking and Countermeasures Research Lab (HCRL), “Can dataset based detection of stealth false data injection attacks in smart grids,” for intrusion detection (otids),” https://ocslab.hksecurity.net/Dataset/ IEEE Systems Journal, vol. 16, no. 2, pp. 2946–2957, 2021. CAN-intrusion-dataset, 2017, accessed: 23 Jul 2024. [180] J. Johnson, T.

Berg, B. Anderson et al., “Review of electric vehicle [196] ——, “Car-hacking dataset for the intrusion detection,” https: charger cybersecurity vulnerabilities, potential impacts, and defenses,” //ocslab.hksecurity.net/Datasets/car-hacking-dataset, 2018, accessed: Energies, vol. 15, no. 11, 2022. 23 Jul 2024. [181] J. Johnson, B.

Anderson, B. Wright et al., “Cybersecurity for electric [197] E. Seo, H.

M. Song, and H. K.

Kim, “Gids: Gan based intrusion vehicle charging infrastructure,” Sandia National Lab.(SNL-NM), detection system for in-vehicle network,” in 2018 16th annual Albuquerque, NM (United States), Tech. Rep., 2022. conference on privacy, security and trust (PST). IEEE, 2018, [182] H.

Lee, S. H. Jeong, and H.

K. Kim, “Otids: A novel intrusion pp. 1–6. detection system for in-vehicle network by using remote frame,” in [198] H. S.

Mavikumbure, V. Cobilean, C. S.

Wickramasinghe et al., 2017 15th Annual Conference on Privacy, Security and Trust (PST). “Physical anomaly detection in ev charging stations: Physics-based IEEE, 2017, pp. 57–5709. vs resnet ae,” in 2023 IEEE 32nd International Symposium on [183] M. E. Verma, M.

D. Iannacone, R. A.

Bridges et al., “Road: The Industrial Electronics (ISIE), 2023, pp. 1–7. real ornl automotive dynamometer controller area network intrusion [199] Z. Liu, Y. Li, Q.

Wang et al., “Tscw-gan based fdias defense for detection dataset (with a comprehensive can ids dataset survey & state-of-charge estimation of battery energy storage systems in smart guide),” ArXiv, vol. abs/2012.14600, 2020. distribution networks,” IEEE Transactions on Industrial Informatics, [184] D. L. Marino, C.

S. Wickramasinghe, K. Amarasinghe et al., “Cyber vol. 20, no. 4, pp. 5048–5059, 2024. and physical anomaly detection in smart-grids,” in 2019 Resilience [200] M.

Li, M. Lenzen, F. Keck et al., “Gis-based probabilistic modeling Week (RWS), vol. 1.

IEEE, 2019, pp. 187–193. of bev charging load for australia,” IEEE Transactions on Smart [185] California ISO, “Open access same-time information system (oasis),” Grid, vol. 10, no. 4, pp. 3525–3534, 2018. http://oasis.caiso.com/mrioasis/logon.do, 2024, accessed: 08 May [201] Y. Liu, W. Liu, S.

Gao et al., “Fast charging demand forecasting 2024. based on the intelligent sensing system of dynamic vehicle under evs- [186] V. Mnih, K. Kavukcuoglu, D.

Silver et al., “Human-level control traffic-distribution coupling,” Energy Reports, vol. 8, pp. 1218–1226, through deep reinforcement learning,” nature, vol. 518, no. 7540, 2022. pp. 529–533, 2015. [202] M. B. Arias, M.

Kim, and S. Bae, “Prediction of electric vehicle [187] T. Lillicrap, “Continuous control with deep reinforcement learning,” charging-power demand in realistic urban traffic networks,” Applied arXiv:1509.02971, 2015. energy, vol. 195, pp. 738–753, 2017. [188] J.

Schulman, F. Wolski, P. Dhariwal et al., “Proximal policy [203] S.

Powell, G. V. Cezar, E.

Apostolaki-Iosifidou et al., “Large-scale optimization algorithms,” arXiv:1707.06347, 2017. scenarios of electric vehicle charging with a data-driven model of [189] T. Haarnoja, A. Zhou, P.

Abbeel et al., “Soft actor-critic: Off-policy control,” Energy, vol. 248, p. 123592, 2022. maximum entropy deep reinforcement learning with a stochastic [204] D. A. R. de Jesús, P.

Mandal, T. Senjyu et al., “Unsupervised hybrid actor,” in International conference on machine learning. PMLR, deep generative models for photovoltaic synthetic data generation,” 2018, pp. 1861–1870. in 2021 IEEE Power & Energy Society General Meeting (PESGM). [190] I.

J. Goodfellow, J. Shlens, and C.

Szegedy, “Explaining and IEEE, 2021, pp. 1–5. harnessing adversarial examples,” arXiv:1412.6572, 2014. [205] J. Yoon, D. Jarrett, and M.

Van der Schaar, “Time-series generative [191] A. Omara and B. Kantarci, “Generative adversarial networks to secure adversarial networks,” Advances in neural information processing vehicle-to-microgrid services,” in 2023 IEEE Virtual Conference on systems, vol. 32, 2019.

Communications (VCC), 2023, pp. 151–156. [206] Shanghai New Energy Electric Vehicle Monitoring Center, “The [192] P. Huber, M. Ott, M.

Friedli et al., “Residential power traces for public data open platform for new energy vehicles,” https:// five houses: the ihomelab rapt dataset,” https://doi.org/10.5281/ labs.shevdc.com/home, 2024, accessed: 13 Jun 2024. zenodo.3581895, 18 Dec 2019, accessed: 08 Jun 2024. [207] City of Boulder government, “City of boulder open data,” https: [193] A. Omara and B. Kantarci, “On the impact of data integrity attacks //open-data.bouldercolorado.gov/, 2021, accessed: 27 May 2024. on vehicle-to-microgrid services,” in 2021 IEEE 26th International [208] Elia Group, “Transparency on grid data,” http://www.elia.be/en/grid- Workshop on Computer Aided Modeling and Design of Communica- data, 2024, accessed: 10 Jun 2024. tion Links and Networks (CAMAD).

IEEE, 2021, pp. 1–7. [209] Customer-Led Network Revolution (CLNR), “What was driving [194] Electric Reliability Council of Texas (ERCOT), “Backcasted (actual) the project?” http://www.networkrevolution.co.uk/the-project/what- load profiles - historical,” https://www.ercot.com/mktinfo/loadprofile/ were-the-drivers/, 2024, accessed: 21 Jul 2024. alp, 2024, accessed: 12 Jun 2024.